Multimodal sentiment analysis (MSA) leverages heterogeneous data sources to interpret the complex nature of human sentiments. Despite significant progress in multimodal architecture design, the field lacks comprehensive regularization methods. This paper introduces PowMix, a versatile embedding space regularizer that builds upon the strengths of unimodal mixing-based regularization approaches and introduces novel algorithmic components that are specifically tailored to multimodal tasks. PowMix is integrated before the fusion stage of multimodal architectures and facilitates intra-modal mixing, such as mixing text with text, to act as a regularizer. PowMix consists of five components: 1) a varying number of generated mixed examples, 2) mixing factor reweighting, 3) anisotropic mixing, 4) dynamic mixing, and 5) cross-modal label mixing. Extensive experimentation across benchmark MSA datasets and a broad spectrum of diverse architectural designs demonstrate the efficacy of PowMix, as evidenced by consistent performance improvements over baselines and existing mixing methods. An in-depth ablation study highlights the critical contribution of each PowMix component and how they synergistically enhance performance. Furthermore, algorithmic analysis demonstrates how PowMix behaves in different scenarios, particularly comparing early versus late fusion architectures. Notably, PowMix enhances overall performance without sacrificing model robustness or magnifying text dominance. It also retains its strong performance in situations of limited data. Our findings position PowMix as a promising versatile regularization strategy for MSA.

Methods based on class activation maps (CAM) provide a simple mechanism to interpret predictions of convolutional neural networks by using linear combinations of feature maps as saliency maps. By contrast, masking-based methods optimize a saliency map directly in the image space or learn it by training another network on additional data.

In this work we introduce Opti-CAM, combining ideas from CAM-based and masking-based approaches. Our saliency map is a linear combination of feature maps, where weights are optimized per image such that the logit of the masked image for a given class is maximized. We also fix a fundamental flaw in two of the most common evaluation metrics of attribution methods. On several datasets, Opti-CAM largely outperforms other CAM-based approaches according to the most relevant classification metrics. We provide empirical evidence supporting that localization and classifier interpretability are not necessarily aligned.

Weakly-supervised object detection attempts to limit the amount of supervision by dispensing the need for bounding boxes, but still assumes image-level labels on the entire training set. In this work, we study the problem of training an object detector from one or few images with image-level labels and a larger set of completely unlabeled images. This is an extreme case of semi-supervised learning where the labeled data are not enough to bootstrap the learning of a detector. Our solution is to train a weakly-supervised student detector model from image-level pseudo-labels generated on the unlabeled set by a teacher classifier model, bootstrapped by region-level similarities to labeled images. Building upon the recent representative weakly-supervised pipeline PCL, our method can use more unlabeled images to achieve performance competitive or superior to many recent weakly-supervised detection solutions.

Adversarial examples of deep neural networks are receiving ever increasing attention because they help in understanding and reducing the sensitivity to their input. This is natural given the increasing applications of deep neural networks in our everyday lives. When white-box attacks are almost always successful, it is typically only the distortion of the perturbations that matters in their evaluation. In this work, we argue that speed is important as well, especially when considering that fast attacks are required by adversarial training. Given more time, iterative methods can always find better solutions. We investigate this speed-distortion trade-off in some depth and introduce a new attack called boundary projection (BP) that improves upon existing methods by a large margin. Our key idea is that the classification boundary is a manifold in the image space: we therefore quickly reach the boundary and then optimize distortion on this manifold.

This paper investigates the visual quality of the adversarial examples. Recent papers propose to smooth the perturbations to get rid of high frequency artefacts. In this work, smoothing has a different meaning as it perceptually shapes the perturbation according to the visual content of the image to be attacked. The perturbation becomes locally smooth on the flat areas of the input image, but it may be noisy on its textured areas and sharp across its edges.

This operation relies on Laplacian smoothing, well-known in graph signal processing, which we integrate in the attack pipeline. We benchmark several attacks with and without smoothing under a white-box scenario and evaluate their transferability. Despite the additional constraint of smoothness, our attack has the same probability of success at lower distortion.

Severe background clutter is challenging in many computer vision tasks, including large-scale image retrieval. Global descriptors, that are popular due to their memory and search efficiency, are especially prone to corruption by such a clutter. Eliminating the impact of the clutter on the image descriptor increases the chance of retrieving relevant images and prevents topic drift due to actually retrieving the clutter in the case of query expansion. In this work, we propose a novel salient region detection method. It captures, in an unsupervised manner, patterns that are both discriminative and common in the dataset. Saliency is based on a centrality measure of a nearest neighbor graph constructed from regional CNN representations of dataset images. The proposed method exploits recent CNN architectures trained for object retrieval to construct the image representation from the salient regions. We improve particular object retrieval on challenging datasets containing small objects.

We present a simple computational model for planar shape decomposition that naturally captures most of the rules and salience measures suggested by psychophysical studies, including the minima and short-cut rules, convexity, and symmetry. It is based on a medial axis representation in ways that have not been explored before and sheds more light into the connection between existing rules like minima and convexity. In particular, vertices of the exterior medial axis directly provide the position and extent of negative minima of curvature, while a traversal of the interior medial axis directly provides a small set of candidate endpoints for part-cuts. The final selection follows a prioritized processing of candidate part-cuts according to a local convexity rule that can incorporate arbitrary salience measures. Neither global optimization nor differentiation is involved. We provide qualitative and quantitative evaluation and comparisons on ground-truth data from psychophysical experiments. With our single computational model, we outperform even an ensemble method on several other competing models.

Local image features are routinely used in state-of-the-art methods to solve many computer vision problems like image retrieval, classification, or 3D registration. As the applications become more complex, the research for better visual features is still active. In this paper we present a feature detector that exploits the inherent geometry of sampled image edges using α-shapes. We propose a novel edge sampling scheme that exploits local shape and investigate different triangulations of sampled points. We also introduce a novel approach to represent the anisotropy in a triangulation along with different feature selection methods. Our detector provides a small number of distinctive features that is ideal for large scale applications, while achieving competitive performance in a series of matching and retrieval experiments.

This paper considers a family of metrics to compare images based on their local descriptors. It encompasses the VLAD descriptor and matching techniques such as Hamming Embedding. Making the bridge between these approaches leads us to propose a match kernel that takes the best of existing techniques by combining an aggregation procedure with a selective match kernel. The representation underpinning this kernel is approximated, providing a large scale image search both precise and scalable, as shown by our experiments on several benchmarks.

We show that the same aggregation procedure, originally applied per image, can effectively operate on groups of similar features found across multiple images. This method implicitly performs feature set augmentation, while enjoying savings in memory requirements at the same time. Finally, the proposed method is shown effective for place recognition, outperforming state of the art methods on a large scale landmark recognition benchmark.

Local feature detection has been an essential part of many methods for computer vision applications like large scale image retrieval, object detection, or tracking. Recently, structure-guided feature detectors have been proposed, exploiting image edges to accurately capture local shape. Among them, the WαSH detector [Varytimidis et al., 2012] starts from sampling binary edges and exploits α-shapes, a computational geometry representation that describes local shape in different scales. In this work, we propose a novel image sampling method, based on dithering smooth image functions other than intensity. Samples are extracted on image contours representing the underlying shapes, with sampling density determined by image functions like the gradient or Hessian response, rather than being fixed. We thoroughly evaluate the parameters of the method, and achieve state-of-the-art performance on a series of matching and retrieval experiments.

We present a new approach to image indexing and retrieval, which integrates appearance with global image geometry in the indexing process, while enjoying robustness against viewpoint change, photometric variations, occlusion, and background clutter. We exploit shape parameters of local features to estimate image alignment via a single correspondence. Then, for each feature, we construct a sparse spatial map of all remaining features, encoding their normalized position and appearance, typically vector quantized to visual word. An image is represented by a collection of such feature maps and RANSAC-like matching is reduced to a number of set intersections. The required index space is still quadratic in the number of features. To make it linear, we propose a novel feature selection model tailored to our feature map representation, replacing our earlier hashing approach. The resulting index space is comparable to baseline bag-of-words, scaling up to one million images while outperforming the state of the art on three publicly available datasets. To our knowledge, this is the first geometry indexing method to dispense with spatial verification at this scale, bringing query times down to milliseconds.

Exploiting local feature shape has made geometry indexing possible, but at a high cost of index space, while a sequential spatial verification and re-ranking stage is still indispensable for large scale image retrieval. In this work we investigate an accelerated approach for the latter problem. We develop a simple spatial matching model inspired by Hough voting in the transformation space, where votes arise from single feature correspondences. Using a histogram pyramid, we effectively compute pair-wise affinities of correspondences without ever enumerating all pairs. Our Hough pyramid matching algorithm is linear in the number of correspondences and allows for multiple matching surfaces or non-rigid objects under one-to-one mapping. We achieve re-ranking one order of magnitude more images at the same query time with superior performance compared to state of the art methods, while requiring the same index space. We show that soft assignment is compatible with this matching scheme, preserving one-to-one mapping and further increasing performance.

Multimodal streams of sensory information are naturally parsed and integrated by humans using signal-level feature extraction and higher-level cognitive processes. Detection of attention-invoking audiovisual segments is formulated in this work on the basis of saliency models for the audio, visual and textual information conveyed in a video stream. Aural or auditory saliency is assessed by cues that quantify multifrequency waveform modulations, extracted through nonlinear operators and energy tracking. Visual saliency is measured through a spatiotemporal attention model driven by intensity, color and orientation. Textual or linguistic saliency is extracted from part-of-speech tagging on the subtitles information available with most movie distributions. The individual saliency streams, obtained from modality-depended cues, are integrated in a multimodal saliency curve, modeling the time-varying perceptual importance of the composite video stream and signifying prevailing sensory events. The multimodal saliency representation forms the basis of a generic, bottom-up video summarization algorithm. Different fusion schemes are evaluated on a movie database of multimodal saliency annotations with comparative results provided across modalities. The produced summaries, based on low-level features and content-independent fusion and selection, are of subjectively high aesthetic and informative quality.

Although the mechanisms of human visual understanding remain partially unclear, computational models inspired by existing knowledge on human vision have emerged and applied to several fields. In this paper, we propose a novel method to compute visual saliency from video sequences by counting in the actual spatiotemporal nature of the video. The visual input is represented by a volume in space–time and decomposed into a set of feature volumes in multiple resolutions. Feature competition is used to produce a saliency distribution of the input implemented by constrained minimization. The proposed constraints are inspired by and associated with the Gestalt laws. There are a number of contributions in this approach, namely extending existing visual feature models to a volumetric representation, allowing competition across features, scales and voxels, and formulating constraints in accordance with perceptual principles. The resulting saliency volume is used to detect prominent spatiotemporal regions and consequently applied to action recognition and perceptually salient event detection in video sequences. Comparisons against established methods on public datasets are given and reveal the potential of the proposed model. The experiments include three action recognition scenarios and salient temporal segment detection in a movie database annotated by humans.

New applications are emerging every day exploiting the huge data volume in community photo collections. Most focus on popular subsets, e.g. images containing landmarks or associated to Wikipedia articles. In this work we are concerned with the problem of accurately finding the location where a photo is taken without needing any metadata, that is, solely by its visual content. We also recognize landmarks where applicable, automatically linking to Wikipedia. We show that the time is right for automating the geo-tagging process, and we show how this can work at large scale. In doing so, we do exploit redundancy of content in popular locations - but unlike most existing solutions, we do not restrict to landmarks. In other words, we can compactly represent the visual content of all thousands of images depicting e.g. the Parthenon and still retrieve any single, isolated, non-landmark image like a house or a graffiti on a wall. Starting from an existing, geo-tagged dataset, we cluster images into sets of different views of the same scene. This is a very efficient, scalable, and fully automated mining process. We then align all views in a set to one reference image and construct a 2D scene map. Our indexing scheme operates directly on scene maps. We evaluate our solution on a challenging one million urban image dataset and provide public access to our service through our application, VIRaL.

Computer vision applications often need to process only a representative part of the visual input rather than the whole image/sequence. Considerable research has been carried out into salient region detection methods based either on models emulating human visual attention (VA) mechanisms or on computational approximations. Most of the proposed methods are bottom-up and their major goal is to filter out redundant visual information. In this paper, we propose and elaborate on a saliency detection model that treats a video sequence as a spatiotemporal volume and generates a local saliency measure for each visual unit (voxel). This computation involves an optimization process incorporating inter- and intra-feature competition at the voxel level. Perceptual decomposition of the input, spatiotemporal center-surround interactions and the integration of heterogeneous feature conspicuity values are described and an experimental framework for video classification is set up. This framework consists of a series of experiments that shows the effect of saliency in classification performance and let us draw conclusions on how well the detected salient regions represent the visual input. A comparison is attempted that shows the potential of the proposed method.

In this paper we investigate detection of high-level concepts in multimedia content through an integrated approach of visual thesaurus analysis and visual context. In the former, detection is based on model vectors that represent image composition in terms of region types, obtained through clustering over a large data set. The latter deals with two aspects, namely high-level concepts and region types of the thesaurus, employing a model of a priori specified semantic relations among concepts and automatically extracted topological relations among region types; thus it combines both conceptual and topological context. A set of algorithms is presented, which modify either the confidence values of detected concepts, or the model vectors based on which detection is performed. Visual context exploitation is evaluated on TRECVID and Corel data sets and compared to a number of related visual thesaurus approaches.

This paper presents a video analysis approach based on concept detection and keyframe extraction employing a visual thesaurus representation. Color and texture descriptors are extracted from coarse regions of each frame and a visual thesaurus is constructed after clustering regions. The clusters, called region types, are used as basis for representing local material information through the construction of a model vector for each frame, which reflects the composition of the image in terms of region types. Model vector representation is used for keyframe selection either in each video shot or across an entire sequence. The selection process ensures that all region types are represented. A number of high-level concept detectors is then trained using global annotation and Latent Semantic Analysis is applied. To enhance detection performance per shot, detection is employed on the selected keyframes of each shot, and a framework is proposed for working on very large data sets.

In this paper we present a framework for unified, personalized access to heterogeneous multimedia content in distributed repositories. Focusing on semantic analysis of multimedia documents, metadata, user queries and user profiles, it contributes to the bridging of the gap between the semantic nature of user queries and raw multimedia documents. The proposed approach utilizes as input visual content analysis results, as well as analyzes and exploits associated textual annotation, in order to extract the underlying semantics, construct a semantic index and classify documents to topics, based on a unified knowledge and semantics representation model. It may then accept user queries, and, carrying out semantic interpretation and expansion, retrieve documents from the index and rank them according to user preferences, similarly to text retrieval. All processes are based on a novel semantic processing methodology, employing fuzzy algebra and principles of taxonomic knowledge representation. Part I of this work presented in this paper deals with data and knowledge models, manipulation of multimedia content annotations and semantic indexing, while Part II will continue on the use of the extracted semantic information for personalized retrieval.

Context modeling has been long acknowledged as a key aspect in a wide variety of problem domains. In this paper we focus on the combination of contextualization and personalization methods to improve the performance of personalized information retrieval. The key aspects in our proposed approach are a) the explicit distinction between historic user context and live user context, b) the use of ontology-driven representations of the domain of discourse, as a common, enriched representational ground for content meaning, user interests, and contextual conditions, enabling the definition of effective means to relate the three of them, and c) the introduction of fuzzy representations as an instrument to properly handle the uncertainty and imprecision involved in the automatic interpretation of meanings, user attention, and user wishes. Based on a formal grounding at the representational level, we propose methods for the automatic extraction of persistent semantic user preferences, and live, ad-hoc user interests, which are combined in order to improve the accuracy and reliability of personalization for retrieval.

In this paper we present a framework for simultaneous image segmentation and object labeling leading to automatic image annotation. Focusing on semantic analysis of images, it contributes to knowledge-assisted multimedia analysis and the bridging of the gap between its semantics and low level visual features. The proposed framework operates at semantic level using possible semantic labels, formally defined as fuzzy sets, to make decisions on handling image regions instead of visual features used traditionally. In order to stress its independence of a specific image segmentation approach we have modified two well known region growing algorithms, i.e. watershed and recursive shortest spanning tree, and compared them with their traditional counterparts. Additionally, a visual context representation and analysis approach is presented, blending global knowledge in interpreting each object locally. Contextual information is based on a novel semantic processing methodology, employing fuzzy algebra and ontological taxonomic knowledge representation. In this process, utilization of contextual knowledge re-adjusts semantic region growing labeling results appropriately, by means of fine-tuning the membership degrees of detected concepts. The performance of the overall methodology is demonstrated on a real-life still image dataset from two popular domains.

Personalized content retrieval aims at improving the retrieval process by taking into account the particular interests of individual users. However, not all user preferences are relevant in all situations. It is well known that human preferences are complex, multiple, heterogeneous, changing, even contradictory, and should be understood in context with the user goals and tasks at hand. In this paper we propose a method to build a dynamic representation of the semantic context of ongoing retrieval tasks, which is used to activate different subsets of user interests at runtime, in a way that out–of-context preferences are discarded. Our approach is based on an ontology-driven representation of the domain of discourse, providing enriched descriptions of the semantics involved in retrieval actions and preferences, and enabling the definition of effective means to relate preferences and context.

A video analysis framework based on spatiotemporal saliency calculation is presented. We propose a novel scheme for generating saliency in video sequences by taking into account both the spatial extent and dynamic evolution of regions. Towards this goal we extend a common image-oriented computational model of saliency-based visual attention to handle spatiotemporal analysis of video in a volumetric framework. The main claim is that attention acts as an efficient preprocessing step of a video sequence in order to obtain a compact representation of its content in the form of salient events/objects. The model has been implemented and qualitative as well as quantitative examples illustrating its performance are shown.

In this article, an approach to semantic image analysis is presented. Under the proposed approach, ontologies are used to capture general, spatial, and contextual knowledge of a domain, and a genetic algorithm is applied to realize the final annotation. The employed domain knowledge considers high-level information in terms of the concepts of interest of the examined domain, contextual information in the form of fuzzy ontological relations, as well as low-level information in terms of prototypical low-level visual descriptors. To account for the inherent ambiguity in visual information, uncertainty has been introduced in the spatial relations definition. First, an initial hypothesis set of graded annotations is produced for each image region, and then context is exploited to update appropriately the estimated degrees of confidence. Finally, a genetic algorithm is applied to decide the most plausible annotation by utilizing the visual and the spatial concepts definitions included in the domain ontology. Experiments with a collection of photographs belonging to two different domains demonstrate the performance of the proposed approach.

Knowledge representation and annotation of multimedia documents typically have been pursued in two different directions. Previous approaches have focused either on low level descriptors, such as dominant color, or on the semantic content dimension and corresponding manual annotations, such as person or vehicle. In this paper, we present a knowledge infrastructure and a experimentation platform for semantic annotation to bridge the two directions. Ontologies are being extended and enriched to include low-level audiovisual features and descriptors. Additionally, we present a tool that allows for linking low-level MPEG-7 visual descriptions to ontologies and annotations. This way we construct ontologies that include prototypical instances of high-level domain concepts together with a formal specification of the corresponding visual descriptors. This infrastructure is exploited by a knowledge-assisted analysis framework that may handle problems like segmentation, tracking, feature extraction and matching in order to classify scenes, identify and label objects, thus automatically create the associated semantic metadata.

The property of transitivity is one of the most important for fuzzy binary relations, especially in the cases when they are used for the representation of real life similarity or ordering information. As far as the algorithmic part of the actual calculation of the transitive closure of such relations is concerned, works in the literature mainly focus on crisp symmetric relations, paying little attention to the case of general fuzzy binary relations. Most works that deal with the algorithmic part of the transitive closure of fuzzy relations only focus on the case of max-min transitivity, disregarding other types of transitivity. In this paper, after formalizing the notion of sparseness and providing a representation model for sparse relations that displays both computational and storage merits, we propose an algorithm for the incremental update of fuzzy sup-t transitive relations. The incremental transitive update (ITU) algorithm achieves the re-establishment of transitivity when an already transitive relation is only locally disturbed. Based on this algorithm, we propose an extension to handle the sup-t transitive closure of any fuzzy binary relation, through a novel incremental transitive closure (ITC) algorithm. The ITU and ITC algorithms can be applied on any fuzzy binary relation and t-norm; properties such as reflexivity, symmetricity and idempotency are not a requirement. Under the specified assumptions for the average sparse relation, both of the proposed algorithms have considerably smaller computational complexity than the conventional approach; this is both established theoretically and verified via appropriate computing experiments.

During the last few years numerous multimedia archives have made extensive use of digitized storage and annotation technologies. Still, the development of single points of access, providing common and uniform access to their data, despite the efforts and accomplishments of standardization organizations, has remained an open issue, as it involves the integration of various large scale heterogeneous and heterolingual systems. In this paper, we describe a mediator system that achieves architectural integration through an extended 3-tier architecture and content integration through semantic modeling. The described system has successfully integrated five multimedia archives, quite different in nature and content from each other, while also providing for easy and scalable inclusion of more archives in the future.

In this paper, an integrated information system is presented that offers enhanced search and retrieval capabilities to users of heterogeneous digital audiovisual (a/v) archives. This innovative system exploits the advances in handling a/v content and related metadata, as introduced by MPEG-4 and worked out by MPEG-7, to offer advanced services characterized by the tri-fold "semantic phrasing of the request (query)", "unified handling" and "personalized response". The proposed system is targeting the intelligent extraction of semantic information from a/v and text related data taking into account the nature of the queries that users my issue, and the context determined by user profiles. It also provides a personalization process of the response in order to provide end-users with desired information. From a technical point of view, the FAETHON system plays the role of an intermediate access server residing between the end users and multiple heterogeneous audiovisual archives organized according to the new MPEG standards.

Pictures and video sequences showing human faces are of high importance in content-based retrieval systems, and consequently face detection has been established as an important tool in the framework of many multimedia applications like indexing, scene classification and news summarisation. In this work, we combine skin colour and shape features with template matching in an efficient way for the purpose of facial image indexing. We propose an adaptive two-dimensional Gaussian model of the skin colour distribution whose parameters are re-estimated based on the current image or frame, reducing generalisation problems. Masked areas obtained from skin colour detection are processed using morphological tools and assessed using global shape features. The verification stage is based on a template matching variation providing robust detection. Facial images and video sequences are indexed according to the number of included faces, their average colour components and their scale, leading to new types of content-based retrieval criteria in query-by-example frameworks. Experimental results have shown that the proposed implementation combines efficiency, robustness and speed, and could be easily embedded in generic visual information retrieval systems or video databases.

A novel method for two-dimensional curve normalization with respect to affine transformations is presented in this paper, which allows an affine-invariant curve representation to be obtained without any actual loss of information on the original curve. It can be applied as a preprocessing step to any shape representation, classification, recognition, or retrieval technique, since it effectively decouples the problem of affine-invariant description from feature extraction and pattern matching. Curves estimated from object contours are first modeled by cubic B-splines and then normalized in several steps in order to eliminate translation, scaling, skew, starting point, rotation, and reflection transformations, based on a combination of curve features including moments and Fourier descriptors.

In this paper, a fuzzy representation of visual content is proposed, which is useful for the new emerging multimedia applications, such as content-based image indexing and retrieval, video browsing and summarization. In particular, a multidimensional fuzzy histogram is constructed for each video frame based on a collection of appropriate features, extracted using video sequence analysis techniques. This approach is then applied both for video summarization, in the context of a content-based sampling algorithm, and for content-based indexing and retrieval. In the first case, video summarization is accomplished by discarding shots or frames of similar visual content so that only a small but meaningful amount of information is retained (key-frames). In the second case, a content-based retrieval scheme is investigated, so that the most similar images to a query are extracted. Experimental results and comparison with other known methods are presented to indicate the good performance of the proposed scheme on real-life video recordings.

An efficient technique for summarization of stereoscopic video sequences is presented in this paper, which extracts a small but meaningful set of video frames using a content-based sampling algorithm. The proposed video-content representation provides the capability of browsing digital stereoscopic video sequences and performing more efficient content-based queries and indexing. Each stereoscopic video sequence is first partitioned into shots by applying a shot-cut detection algorithm so that frames (or stereo pairs) of similar visual characteristics are gathered together. Each shot is then analyzed using stereo-imaging techniques, and the disparity field, occluded areas, and depth map are estimated. A multiresolution implementation of the Recursive Shortest Spanning Tree (RSST) algorithm is applied for color and depth segmentation, while fusion of color and depth segments is employed for reliable video object extraction. In particular, color segments are projected onto depth segments so that video objects on the same depth plane are retained, while at the same time accurate object boundaries are extracted. Feature vectors are then constructed using multidimensional fuzzy classification of segment features including size, location, color, and depth. Shot selection is accomplished by clustering similar shots based on the generalized Lloyd-Max algorithm, while for a given shot, key frames are extracted using an optimization method for locating frames of minimally correlated feature vectors. For efficient implementation of the latter method, a genetic algorithm is used. Experimental results are presented, which indicate the reliable performance of the proposed scheme on real-life stereoscopic video sequences.

A video content representation framework is proposed in this paper for extracting limited, but meaningful, information of video data, directly from the MPEG compressed domain. A hierarchical color and motion segmentation scheme is applied to each video shot, transforming the frame-based representation to a feature-based one. The scheme is based on a multiresolution implementation of the recursive shortest spanning tree (RSST) algorithm. Then, all segment features are gathered together using a fuzzy multidimensional histogram to reduce the possibility of classifying similar segments to different classes. Extraction of several key frames is performed for each shot in a content-based rate-sampling framework. Two approaches are examined for key frame extraction. The first is based on examination of the temporal variation of the feature vector trajectory; the second is based on minimization of a cross-correlation criterion of the video frames. For efficient implementation of the latter approach, a logarithmic search (along with a stochastic version) and a genetic algorithm are proposed. Experimental results are presented which illustrate the performance of the proposed techniques, using synthetic and real life MPEG video sequences.

This work addresses composed image retrieval in the context of domain conversion, where the content of the query image is to be retrieved in a domain given by the query text. We show that a strong generic vision-language model already provides sufficient descriptive power and no further learning is necessary. The query image is mapped to the input text space by textual inversion. In contrast to the common practice of inverting to the continuous space of textual tokens, we opt for inversion into the discrete space of words using a text vocabulary. This distinction is empirically validated and proven to be a pivotal factor. Through inversion, we represent the image by soft assignment to the vocabulary. Such a text description is made more robust by engaging a set of images visually similar to the query image. Images are retrieved via a weighted ensemble of text queries, each composed of one of the words assigned to the query image and the domain query text. Our method outperforms prior art by a large margin on standard as well as on newly introduced benchmarks.

Explanations obtained from transformer-based architectures, in the form of raw attention, can be seen as a class agnostic saliency map. Additionally, attention-based pooling serves as a form of masking in feature space. Motivated by this observation, we design an attention-based pooling mechanism intended to replace global average pooling during inference. This mechanism, called Cross Attention Stream (CA-Stream), comprises a stream of cross attention blocks interacting with features at different network levels. CA-Stream enhances interpretability properties in existing image recognition models, while preserving their recognition properties.

The surge in data volume within the field of remote sensing has necessitated efficient methods for retrieving relevant information from extensive image archives. Conventional unimodal queries, whether visual or textual, are insufficient and restrictive. To address this limitation, we introduce the task of composed image retrieval in remote sensing, allowing users to combine query images with a textual part that modifies attributes such as color, texture, context, or more, thereby enhancing the expressivity of the query.

We demonstrate that a vision-language model possesses sufficient descriptive power and, when coupled with the proposed fusion method, eliminates the necessity for further learning. We present a new evaluation benchmark focused on shape, color, density, and quantity modifications. Our work not only sets the state-of-the-art for this task, but also serves as a foundational step in addressing a gap in the field of remote sensing image retrieval.

How important is it for training and evaluation sets to not have class overlap in image retrieval? We revisit Google Landmarks v2 clean, the most popular training set, by identifying and removing class overlap with Revisited Oxford and Paris, the most popular evaluation set. By comparing the original and the new $\mathcal{R}$GLDv2-clean on a benchmark of reproduced state-of-the-art methods, our findings are striking. Not only is there a dramatic drop in performance, but it is inconsistent across methods, changing the ranking.

What does it take to focus on objects or interest and ignore background clutter when indexing? Do we need to train an object detector and the representation separately? Do we need location supervision? We introduce Single-stage Detect-to-Retrieve (CiDeR), an end-to-end, single-stage pipeline to detect objects of interest and extract a global image representation. We outperform previous state-of-the-art on both existing training sets and the new $\mathcal{R}$GLDv2-clean. Our dataset is available at https://github.com/dealicious-inc/RGLDv2-clean.

Self-supervised learning has unlocked the potential of scaling up pretraining to billions of images, since annotation is unnecessary. But are we making the best use of data? How more economical can we be? In this work, we attempt to answer this question by making two contributions. First, we investigate first-person videos and introduce a "Walking Tours" dataset. These videos are high-resolution, hours-long, captured in a single uninterrupted take, depicting a large number of objects and actions with natural scene transitions. They are unlabeled and uncurated, thus realistic for self-supervision and comparable with human learning.

Second, we introduce a novel self-supervised image pretraining method tailored for learning from continuous videos. Existing methods typically adapt image-based pretraining approaches to incorporate more frames. Instead, we advocate a "tracking to learn to recognize" approach. Our method called DoRA, leads to attention maps that DiscOver and tRAck objects over time in an end-to-end manner, using transformer cross-attention. We derive multiple views from the tracks and use them in a classical self-supervised distillation loss. Using our novel approach, a single Walking Tours video remarkably becomes a strong competitor to ImageNet for several image and video downstream tasks.

This paper studies interpretability of convolutional networks by means of saliency maps. Most approaches based on Class Activation Maps (CAM) combine information from fully connected layers and gradient through variants of backpropagation. However, it is well understood that gradients are noisy and alternatives like guided backpropagation have been proposed to obtain better visualization at inference. In this work, we present a novel training approach to improve the quality of gradients for interpretability. In particular, we introduce a regularization loss such that the gradient with respect to the input image obtained by standard backpropagation is similar to the gradient obtained by guided backpropagation. We find that the resulting gradient is qualitatively less noisy and improves quantitatively the interpretability properties of different networks, using several interpretability methods.

Transductive few-shot learning algorithms have showed substantially superior performance over their inductive counterparts by leveraging the unlabeled queries at inference. However, the vast majority of transductive methods are evaluated on perfectly class-balanced benchmarks. It has been shown that they undergo remarkable drop in performance under a more realistic, imbalanced setting.

To this end, we propose a novel algorithm to address imbalanced transductive few-shot learning, named Adaptive Manifold. Our algorithm exploits the underlying manifold of the labeled examples and unlabeled queries by using manifold similarity to predict the class probability distribution of every query. It is parameterized by one centroid per class and a set of manifold parameters that determine the manifold. All parameters are optimized by minimizing a loss function that can be tuned towards class-balanced or imbalanced distributions. The manifold similarity shows substantial improvement over Euclidean distance, especially in the 1-shot setting.

Our algorithm outperforms all other state of the art methods in three benchmark datasets, namely miniImageNet, tieredImageNet and CUB, and two different backbones, namely ResNet-18, WideResNet-28-10. In certain cases, our algorithm outperforms the previous state of the art by as much as 4.2%.

Mixup refers to interpolation-based data augmentation, originally motivated as a way to go beyond empirical risk minimization (ERM). Its extensions mostly focus on the definition of interpolation and the space (input or embedding) where it takes place, while the augmentation process itself is less studied. In most methods, the number of generated examples is limited to the mini-batch size and the number of examples being interpolated is limited to two (pairs), in the input space.

We make progress in this direction by introducing MultiMix, which generates an arbitrarily large number of interpolated examples beyond the mini-batch size, and interpolates the entire mini-batch in the embedding space. Effectively, we sample on the entire convex hull of the mini-batch rather than along linear segments between pairs of examples.

On sequence data we further extend to Dense MultiMix. We densely interpolate features and target labels at each spatial location and also apply the loss densely. To mitigate the lack of dense labels, we inherit labels from examples and weight interpolation factors by attention as a measure of confidence.

Overall, we increase the number of loss terms per mini-batch by orders of magnitude at little additional cost. This is only possible because of interpolating in the embedding space. We empirically show that our solutions yield significant improvement over state-of-the-art mixup methods on four different benchmarks, despite interpolation being only linear. By analyzing the embedding space, we show that the classes are more tightly clustered and uniformly spread over the embedding space, thereby explaining the improved behavior.

Recent vision-language models are driven by large-scale pretrained models. However, adapting pretrained models on limited data presents challenges such as overfitting, catastrophic forgetting, and the cross-modal gap between vision and language. We introduce a parameter-efficient method to address these challenges, combining multimodal prompt learning and a transformer-based mapping network, while keeping the pretrained models frozen. Our experiments on several video question answering benchmarks demonstrate the superiority of our approach in terms of performance and parameter efficiency on both zero-shot and few-shot settings. Our code is available at https://engindeniz.github.io/vitis.

Convolutional networks and vision transformers have different forms of pairwise interactions, pooling across layers and pooling at the end of the network. Does the latter really need to be different? As a by-product of pooling, vision transformers provide spatial attention for free, but this is most often of low quality unless self-supervised, which is not well studied. Is supervision really the problem? In this work, we develop a generic pooling framework and then we formulate a number of existing methods as instantiations. By discussing the properties of each group of methods, we derive SimPool, a simple attention-based pooling mechanism as a replacement of the default one for both convolutional and transformer encoders. We find that, whether supervised or self-supervised, this improves performance on pre-training and downstream tasks and provides attention maps delineating object boundaries in all cases. One could thus call SimPool universal. To our knowledge, we are the first to obtain attention maps in supervised transformers of at least as good quality as self-supervised, without explicit losses or modifying the architecture. Code at: https://github.com/billpsomas/simpool.

Few-shot learning addresses the issue of classifying images using limited labeled data. Exploiting unlabeled data through the use of transductive inference methods such as label propagation has been shown to improve the performance of few-shot learning significantly. Label propagation infers pseudo-labels for unlabeled data by utilizing a constructed graph that exploits the underlying manifold structure of the data. However, a limitation of the existing label propagation approaches is that the positions of all data points are fixed and might be sub-optimal so that the algorithm is not as effective as possible. In this work, we propose a novel algorithm that adapts the feature embeddings of the labeled data by minimizing a differentiable loss function optimizing their positions in the manifold in the process. Our novel algorithm, Adaptive Anchor Label Propagation, outperforms the standard label propagation algorithm by as much as 7% and 2% in the 1-shot and 5-shot settings respectively. We provide experimental results highlighting the merits of our algorithm on four widely used few-shot benchmark datasets, namely miniImageNet, tieredImageNet, CUB and CIFAR-FS and two commonly used backbones, ResNet12 and WideResNet-28-10. The source code can be found at https://github.com/MichalisLazarou/A2LP.

Impressive progress in generative models and implicit representations gave rise to methods that can generate 3D shapes of high quality. However, being able to locally control and edit shapes is another essential property that can unlock several content creation applications. Local control can be achieved with part-aware models, but existing methods require 3D supervision and cannot produce textures. In this work, we devise PartNeRF, a novel part-aware generative model for editable 3D shape synthesis that does not require any explicit 3D supervision. Our model generates objects as a set of locally defined NeRFs, augmented with an affine transformation. This enables several editing operations such as applying transformations on parts, mixing parts from different objects etc. To ensure distinct, manipulable parts we enforce a hard assignment of rays to parts that makes sure that the color of each ray is only determined by a single NeRF. As a result, altering one part does not affect the appearance of the others. Evaluations on various ShapeNet categories demonstrate the ability of our model to generate editable 3D objects of improved fidelity, compared to previous part-based generative approaches that require 3D supervision or models relying on NeRFs.

Vision transformers have achieved remarkable progress in vision tasks such as image classification and detection. However, in instance-level image retrieval, transformers have not yet shown good performance compared to convolutional networks. We propose a number of improvements that make transformers outperform the state of the art for the first time. (1) We show that a hybrid architecture is more effective than plain transformers, by a large margin. (2) We introduce two branches collecting global (classification token) and local (patch tokens) information, from which we form a global image representation. (3) In each branch, we collect multi-layer features from the transformer encoder, corresponding to skip connections across distant layers. (4) We enhance locality of interactions at the deeper layers of the encoder, which is the relative weakness of vision transformers. We train our model on all commonly used training sets and, for the first time, we make fair comparisons separately per training set. In all cases, we outperform previous models based on global representation. Public code is available at https://github.com/dealicious-inc/DToP.

Transformers and masked language modeling are quickly being adopted and explored in computer vision as vision transformers and masked image modeling (MIM). In this work, we argue that image token masking differs from token masking in text, due to the amount and correlation of tokens in an image. In particular, to generate a challenging pretext task for MIM, we advocate a shift from random masking to informed masking. We develop and exhibit this idea in the context of distillation-based MIM, where a teacher transformer encoder generates an attention map, which we use to guide masking for the student.

We thus introduce a novel masking strategy, called attention-guided masking (AttMask), and we demonstrate its effectiveness over random masking for dense distillation-based MIM as well as plain distillation-based self-supervised learning on classification tokens. We confirm that AttMask accelerates the learning process and improves the performance on a variety of downstream tasks. We provide the implementation code at https://github.com/gkakogeorgiou/attmask.

Mixup is a powerful data augmentation method that interpolates between two or more examples in the input or feature space and between the corresponding target labels. Many recent mixup methods focus on cutting and pasting two or more objects into one image, which is more about efficient processing than interpolation. However, how to best interpolate images is not well defined. In this sense, mixup has been connected to autoencoders, because often autoencoders "interpolate well", for instance generating an image that continuously deforms into another.

In this work, we revisit mixup from the interpolation perspective and introduce AlignMix, where we geometrically align two images in the feature space. The correspondences allow us to interpolate between two sets of features, while keeping the locations of one set. Interestingly, this gives rise to a situation where mixup retains mostly the geometry or pose of one image and the texture of the other, connecting it to style transfer. More than that, we show that an autoencoder can still improve representation learning under mixup, without the classifier ever seeing decoded images. AlignMix outperforms state-of-the-art mixup methods on five different benchmarks.

Metric learning involves learning a discriminative representation such that embeddings of similar classes are encouraged to be close, while embeddings of dissimilar classes are pushed far apart. State-of-the-art methods focus mostly on sophisticated loss functions or mining strategies. On the one hand, metric learning losses consider two or more examples at a time. On the other hand, modern data augmentation methods for classification consider two or more examples at a time. The combination of the two ideas is under-studied.

In this work, we aim to bridge this gap and improve representations using mixup, which is a powerful data augmentation approach interpolating two or more examples and corresponding target labels at a time. This task is challenging because, unlike classification, the loss functions used in metric learning are not additive over examples, so the idea of interpolating target labels is not straightforward. To the best of our knowledge, we are the first to investigate mixing both examples and target labels for deep metric learning. We develop a generalized formulation that encompasses existing metric learning loss functions and modify it to accommodate for mixup, introducing Metric Mix, or Metrix. We also introduce a new metric---utilization---to demonstrate that by mixing examples during training, we are exploring areas of the embedding space beyond the training classes, thereby improving representations. To validate the effect of improved representations, we show that mixing inputs, intermediate representations or embeddings along with target labels significantly outperforms state-of-the-art metric learning methods on four benchmark deep metric learning datasets.

We address representation learning for large-scale instance-level image retrieval. Apart from backbone, training pipelines and loss functions, popular approaches have focused on different spatial pooling and attention mechanisms, which are at the core of learning a powerful global image representation. There are different forms of attention according to the interaction of elements of the feature tensor (local and global) and the dimensions where it is applied (spatial and channel). Unfortunately, each study addresses only one or two forms of attention and applies it to different problems like classification, detection or retrieval.

We present global-local attention module (GLAM), which is attached at the end of a backbone network and incorporates all four forms of attention: local and global, spatial and channel. We obtain a new feature tensor and, by spatial pooling, we learn a powerful embedding for image retrieval. Focusing on global descriptors, we provide empirical evidence of the interaction of all forms of attention and improve the state of the art on standard benchmarks.

Few-shot learning addresses the challenge of learning how to address novel tasks given not just limited supervision but limited data as well. An attractive solution is synthetic data generation. However, most such methods are overly sophisticated, focusing on high-quality, realistic data in the input space. It is unclear whether adapting them to the few-shot regime and using them for the downstream task of classification is the right approach. Previous works on synthetic data generation for few-shot classification focus on exploiting complex models, e.g. a Wasserstein GAN with multiple regularizers or a network that transfers latent diversities from known to novel classes.

We follow a different approach and investigate how a simple and straightforward synthetic data generation method can be used effectively. We make two contributions, namely we show that: (1) using a simple loss function is more than enough for training a feature generator in the few-shot setting; and (2) learning to generate tensor features instead of vector features is superior. Extensive experiments on miniImagenet, CUB and CIFAR-FS datasets show that our method sets a new state of the art, outperforming more sophisticated few-shot data augmentation methods. The source code can be found at https://github.com/MichalisLazarou/TFH_fewshot.

Deep Neural Networks (DNNs) are robust against intra-class variability of images, pose variations and random noise, but vulnerable to imperceptible adversarial perturbations that are well-crafted precisely to mislead. While random noise even of relatively large magnitude can hardly affect predictions, adversarial perturbations of very small magnitude can make a classifier fail completely.

To enhance robustness, we introduce a new adversarial defense called patch replacement, which transforms both the input images and their intermediate features at early layers to make adversarial perturbations behave similarly to random noise. We decompose images/features into small patches and quantize them according to a codebook learned from legitimate training images. This maintains the semantic information of legitimate images, while removing as much as possible the effect of adversarial perturbations.

Experiments show that patch replacement improves robustness against both white-box and gray-box attacks, compared with other transformation-based defenses. It has a low computational cost since it does not need training or fine-tuning the network. Importantly, in the white-box scenario, it increases the robustness, while other transformation-based defenses do not.

High-level understanding of stories in video such as movies and TV shows from raw data is extremely challenging. Modern video question answering (VideoQA) systems often use additional human-made sources like plot synopses, scripts, video descriptions or knowledge bases. In this work, we present a new approach to understand the whole story without such external sources. The secret lies in the dialog: unlike any prior work, we treat dialog as a noisy source to be converted into text description via dialog summarization, much like recent methods treat video. The input of each modality is encoded by transformers independently, and a simple fusion method combines all modalities, using soft temporal attention for localization over long inputs. Our model outperforms the state of the art on the KnowIT VQA dataset by a large margin, without using question-specific human annotation or human-made plot summaries. It even outperforms human evaluators who have never watched any whole episode before.

Few-shot learning amounts to learning representations and acquiring knowledge such that novel tasks may be solved with both supervision and data being limited. Improved performance is possible by transductive inference, where the entire test set is available concurrently, and semi-supervised learning, where more unlabeled data is available. These problems are closely related because there is little or no adaptation of the representation in novel tasks.

Focusing on these two settings, we introduce a new algorithm that leverages the manifold structure of the labeled and unlabeled data distribution to predict pseudo-labels, while balancing over classes and using the loss value distribution of a limited-capacity classifier to select the cleanest labels, iterately improving the quality of pseudo-labels. Our solution sets new state of the art results on four benchmark datasets, namely miniImageNet, tieredImageNet, CUB and CIFAR-FS, while being robust over feature space pre-processing and the quantity of available data.

Knowledge transfer from large teacher models to smaller student models has recently been studied for metric learning, focusing on fine-grained classification. In this work, focusing on instance-level image retrieval, we study an asymmetric testing task, where the database is represented by the teacher and queries by the student. Inspired by this task, we introduce asymmetric metric learning, a novel paradigm of using asymmetric representations at training. This acts as a simple combination of knowledge transfer with the original metric learning task.

We systematically evaluate different teacher and student models, metric learning and knowledge transfer loss functions on the new asymmetric testing as well as the standard symmetric testing task, where database and queries are represented by the same model. We find that plain regression is surprisingly effective compared to more complex knowledge transfer mechanisms, working best in asymmetric testing. Interestingly, our asymmetric metric learning approach works best in symmetric testing, allowing the student to even outperform the teacher.

The challenge in few-shot learning is that available data is not enough to capture the underlying distribution. To mitigate this, two emerging directions are (a) using local image representations, essentially multiplying the amount of data by a constant factor, and (b) using more unlabeled data, for instance by transductive inference, jointly on a number of queries. In this work, we bring these two ideas together, introducing local propagation. We treat local image features as independent examples, we build a graph on them and we use it to propagate both the features themselves and the labels, known and unknown. Interestingly, since there is a number of features per image, even a single query gives rise to transductive inference. As a result, we provide a universally safe choice for few-shot inference under both non-transductive and transductive settings, improving accuracy over corresponding methods. This is in contrast to existing solutions, where one needs to choose the method depending on the quantity of available data.

Few-shot learning is often motivated by the ability of humans to learn new tasks from few examples. However, standard few-shot classification benchmarks assume that the representation is learned on a limited amount of base class data, ignoring the amount of prior knowledge that a human may have accumulated before learning new tasks. At the same time, even if a powerful representation is available, it may happen in some domain that base class data are limited or non-existent. This motivates us to study a problem where the representation is obtained from a classifier pre-trained on a large-scale dataset of a different domain, assuming no access to its training process, while the base class data are limited to few examples per class and their role is to adapt the representation to the domain at hand rather than learn from scratch. We adapt the representation in two stages, namely on the few base class data if available and on the even fewer data of new tasks. In doing so, we obtain from the pre-trained classifier a spatial attention map that allows focusing on objects and suppressing background clutter. This is important in the new problem, because when base class data are few, the network cannot learn where to focus implicitly. We also show that a pre-trained network may be easily adapted to novel classes, without meta-learning.

In this work we consider the problem of learning a classifier from noisy labels when a few clean labeled examples are given. The structure of clean and noisy data is modeled by a graph per class and Graph Convolutional Networks (GCN) are used to predict class relevance of noisy examples. For each class, the GCN is treated as a binary classifier, which learns to discriminate clean from noisy examples using a weighted binary cross-entropy loss function. The GCN-inferred "clean" probability is then exploited as a relevance measure. Each noisy example is weighted by its relevance when learning a classifier for the end task. We evaluate our method on an extended version of a few-shot learning problem, where the few clean examples of novel classes are supplemented with additional noisy data. Experimental results show that our GCN-based cleaning process significantly improves the classification accuracy over not cleaning the noisy data, as well as standard few-shot classification where only few clean examples are used.

Active learning typically focuses on training a model on few labeled examples alone, while unlabeled ones are only used for acquisition. In this work we depart from this setting by using both labeled and unlabeled data during model training across active learning cycles. We do so by using unsupervised feature learning at the beginning of the active learning pipeline and semi-supervised learning at every active learning cycle, on all available data. The former has not been investigated before in active learning, while the study of latter in the context of deep learning is scarce and recent findings are not conclusive with respect to its benefit. Our idea is orthogonal to acquisition strategies by using more data, much like ensemble methods use more models. By systematically evaluating on a number of popular acquisition strategies and datasets, we find that the use of unlabeled data during model training brings a spectacular accuracy improvement in image classification, compared to the differences between acquisition strategies. We thus explore smaller label budgets, even one label per class.

Semi-supervised learning is becoming increasingly important because it can combine data carefully labeled by humans with abundant unlabeled data to train deep neural networks. Classic works on semi-supervised learning that have focused on transductive learning have not been fully exploited in the inductive framework followed by modern deep learning. The same holds for the manifold assumption--that similar examples should get the same prediction. In this work, we employ a transductive method that is based on the manifold assumption to make predictions on the entire dataset and use these predictions to generate pseudo-labels for the unlabeled data and train a deep neural network. In doing so, a nearest neighbor graph of the dataset is created based on the embeddings of the same network. Therefore our learning process iterates between these two steps. We improve performance on several datasets especially in the few labels regime and show that our work is complementary to current state of the art.

Training deep neural networks from few examples is a highly challenging and key problem for many computer vision tasks. In this context, we are targeting knowledge transfer from a set with abundant data to other sets with few available examples. We propose two simple and effective solutions: (i) dense classification over feature maps, which for the first time studies local activations in the domain of few-shot learning, and (ii) implanting, that is, attaching new neurons to a previously trained network to learn new, task-specific features. On miniImageNet, we improve the prior state-of-the-art on few-shot classification, i.e., we achieve 62.5%, 79.8% and 83.8% on 5-way 1-shot, 5-shot and 10-shot settings respectively.

We propose a novel method of spatial verification for image retrieval. Initial ranking is based on image descriptors extracted from convolutional neural network activations by global pooling, as in recent state-of-the-art work. However, the same sparse 3D activation tensor is also approximated by a collection of local features. These local features are then robustly matched to approximate the optimal alignment of the tensors. This happens without any network modification, additional layers or training. No local feature detection happens on the original image; no local feature descriptors and no visual vocabulary are needed throughout the whole process.

We experimentally show that the proposed method achieves the state-of-the-art performance on standard benchmarks across different network architectures and different global pooling methods. Advantages of combining efficient nearest neighbor retrieval with global descriptors and spatial verification is even more pronounced by spatially verified diffusion.

State of the art image retrieval performance is achieved with CNN features and manifold ranking using a k-NN similarity graph that is pre-computed off-line. The two most successful existing approaches are temporal filtering, where manifold ranking amounts to solving a sparse linear system online, and spectral filtering, where eigen-decomposition of the adjacency matrix is performed off-line and then manifold ranking amounts to dot-product search online. The former suffers from expensive queries and the latter from significant space overhead. Here we introduce a novel, theoretically well-founded hybrid filtering approach allowing full control of the space-time trade-off between these two extremes. Experimentally, we verify that our hybrid method delivers results on par with the state of the art, with lower memory demands compared to spectral filtering approaches and faster compared to temporal filtering.

In this work we present a novel unsupervised framework for hard training example mining. The only input to the method is a collection of images relevant to the target application and a meaningful initial representation, provided e.g. by pre-trained CNN. Positive examples are distant points on a single manifold, while negative examples are nearby points on different manifolds. Both types of examples are revealed by disagreements between Euclidean and manifold similarities. The discovered examples can be used in training with any discriminative loss.

The method is applied to unsupervised fine-tuning of pre-trained networks for fine-grained classification and particular object retrieval. Our models are on par or are outperforming prior models that are fully or partially supervised.

In this paper we address issues with image retrieval benchmarking on standard and popular Oxford 5k and Paris 6k datasets. In particular, annotation errors, the size of the dataset, and the level of challenge are addressed: new annotation for both datasets is created with an extra attention to the reliability of the ground truth. Three new protocols of varying difficulty are introduced. The protocols allow fair comparison between different methods, including those using a dataset pre-processing stage. For each dataset, 15 new challenging queries are introduced. Finally, a new set of 1M hard, semi-automatically cleaned distractors is selected.

An extensive comparison of the state-of-the-art methods is performed on the new benchmark. Different types of methods are evaluated, ranging from local-feature-based to modern CNN based methods. The best results are achieved by taking the best of the two worlds. Most importantly, image retrieval appears far from being solved.

Despite the success of deep learning on representing images for particular object retrieval, recent studies show that the learned representations still lie on manifolds in a high dimensional space. This makes the Euclidean nearest neighbor search biased for this task. Exploring the manifolds online remains expensive even if a nearest neighbor graph has been computed offline.

This work introduces an explicit embedding reducing manifold search to Euclidean search followed by dot product similarity search. This is equivalent to linear graph filtering of a sparse signal in the frequency domain. To speed up online search, we compute an approximate Fourier basis of the graph offline. We improve the state of art on particular object retrieval datasets including the challenging Instre dataset containing small objects. At a scale of 10^5 images, the offline cost is only a few hours, while query time is comparable to standard similarity search.

Severe background clutter is challenging in many computer vision tasks, including large-scale image retrieval. Global descriptors, that are popular due to their memory and search efficiency, are especially prone to corruption by such a clutter. Eliminating the impact of the clutter on the image descriptor increases the chance of retrieving relevant images and prevents topic drift due to actually retrieving the clutter in the case of query expansion. In this work, we propose a novel salient region detection method. It captures, in an unsupervised manner, patterns that are both discriminative and common in the dataset. Saliency is based on a centrality measure of a nearest neighbor graph constructed from regional CNN representations of dataset images. The descriptors derived from the salient regions improve particular object retrieval, most noticeably in a large collections containing small objects.

Part-based image classification consists in representing categories by small sets of discriminative parts upon which a representation of the images is built. This paper addresses the question of how to automatically learn such parts from a set of labeled training images. We propose to cast the training of parts as a quadratic assignment problem in which optimal correspondences between image regions and parts are automatically learned. The paper analyses different assignment strategies and thoroughly evaluates them on two public datasets: Willow actions and MIT 67 scenes.

Query expansion is a popular method to improve the quality of image retrieval with both conventional and CNN representations. It has been so far limited to global image similarity. This work focuses on diffusion, a mechanism that captures the image manifold in the feature space. The diffusion is carried out on descriptors of overlapping image regions rather than on a global image descriptor like in previous approaches. An efficient off-line stage allows optional reduction in the number of stored regions. In the on-line stage, the proposed handling of unseen queries in the indexing stage removes additional computation to adjust the precomputed data. A novel way to perform diffusion through a sparse linear system solver yields practical query times well below one second. Experimentally, we observe a significant boost in performance of image retrieval with compact CNN descriptors on standard benchmarks, especially when the query object covers only a small part of the image. Small objects have been a common failure case of CNN-based retrieval.

Part-based image classification aims at representing categories by small sets of learned discriminative parts, upon which an image representation is built. Considered as a promising avenue a decade ago, this direction has been neglected since the advent of deep neural networks. In this context, this paper brings two contributions: first, this work proceeds one step further compared to recent part-based models (PBM), focusing on how to learn parts without using any labeled data. Instead of learning a set of parts per class, as generally performed in the PBM literature, the proposed approach both constructs a partition of a given set of images into visually similar groups, and subsequently learns a set of discriminative parts per group in a fully unsupervised fashion. This strategy opens the door to the use of PBM in new applications where labeled data are typically not available, such as instance-based image retrieval. Second, this paper shows that despite the recent success of end-to-end models, explicit part learning can still boost classification performance. We experimentally show that our learned parts can help building efficient image representations, which outperform state-of-the art Deep Convolutional Neural Networks (DCNN) on both classification and retrieval tasks.

Location recognition is commonly treated as visual instance retrieval on "street view" imagery. The dataset items and queries are panoramic views, i.e. groups of images taken at a single location. This work introduces a novel panorama-to-panorama matching process, either by aggregating features of individual images in a group or by explicitly constructing a larger panorama. In either case, multiple views are used as queries. We reach near perfect location recognition on a standard benchmark with only four query views.

We propose a new data-structure, the generalized randomized k-d forest, or k-d GeRaF, for approximate nearest neighbor searching in high dimensions. In particular, we introduce new randomization techniques to specify a set of independently constructed trees where search is performed simultaneously, hence increasing accuracy. We omit backtracking, and we optimize distance computations, thus accelerating queries. We release public domain software GeRaF and we compare it to existing implementations of state-of-the-art methods including BBD-trees, Locality Sensitive Hashing, randomized k-d forests, and product quantization. Experimental results indicate that our method would be the method of choice in dimensions around 1,000, and probably up to 10,000, and pointsets of cardinality up to a few hundred thousands or even one million; this range of inputs is encountered in many critical applications today. For instance, we handle a real dataset of 10^6 images represented in 960 dimensions with a query time of less than 1sec on average and 90% responses being true nearest neighbors.

Large scale duplicate detection, clustering and mining of documents or images has been conventionally treated with seed detection via hashing, followed by seed growing heuristics using fast search. Principled clustering methods, especially kernelized and spectral ones, have higher complexity and are difficult to scale above millions. Under the assumption of documents or images embedded in Euclidean space, we revisit recent advances in approximate k-means variants, and borrow their best ingredients to introduce a new one, inverted-quantized k-means (IQ-means). Key underlying concepts are quantization of data points and multi-index based inverted search from centroids to cells. Its quantization is a form of hashing and analogous to seed detection, while its updates are analogous to seed growing, yet principled in the sense of distortion minimization. We further design a dynamic variant that is able to determine the number of clusters k in a single run at nearly zero additional cost. Combined with powerful deep learned representations, we achieve clustering of a 100 million image collection on a single machine in less than one hour.

We present a very simple computational model for planar shape decomposition that naturally captures most of the rules and salience measures suggested by psychophysical studies, including the minima and short-cut rules, convexity, and symmetry. It is based on a medial axis representation in ways that have not been explored before and sheds more light into the connection between existing rules like minima and convexity. In particular, vertices of the exterior medial axis directly provide the position and extent of negative minima of curvature, while a traversal of the interior medial axis directly provides a small set of candidate endpoints for part-cuts. The final selection follows a simple local convexity rule that can incorporate arbitrary salience measures. Neither global optimization nor differentiation is involved. We provide qualitative and quantitative evaluation and comparisons on ground-truth data from psychophysical experiments.

Recent works show that image comparison based on local descriptors is corrupted by visual bursts, which tend to dominate the image similarity. The existing strategies, like power-law normalization, improve the results by discounting the contribution of visual bursts to the image similarity.

In this paper, we propose to explicitly detect the visual bursts in an image at an early stage. We compare several detection strategies jointly taking into account feature similarity and geometrical quantities. The bursty groups are merged into meta-features, which are used as input to state-of-the-art image search systems such as VLAD or the selective match kernel. Then, we show the interest of using this strategy in an asymmetrical manner, with only the database features being aggregated but not those of the query.

Extensive experiments performed on public benchmarks for visual retrieval show the benefits of our method, which achieves performance on par with the state of the art but with a significantly reduced complexity, thanks to the lower number of features fed to the indexing system.

The recent trend of structure-guided feature detectors, as opposed to blob and corner detectors, has led to a family of methods that exploit image edges to accurately capture local shape. Among them, the WaSH detector combines binary edge sampling with gradient strength and computational geometry representations towards distinctive and repeatable local features. In this work, we provide alternative, variable-density sampling schemes on smooth functions of image intensity based on dithering. These methods are parameter-free and more invariant to geometric transformations than uniform sampling. The resulting detectors compare well to the state-of-the-art, while achieving higher performance in a series of matching and retrieval experiments.

We present a simple vector quantizer that combines low distortion with fast search and apply it to approximate nearest neighbor (ANN) search in high dimensional spaces. Leveraging the very same data structure that is used to provide non-exhaustive search, i.e. inverted lists or a multi-index, the idea is to locally optimize an individual product quantizer (PQ) per cell and use it to encode residuals. Local optimization is over rotation and space decomposition; interestingly, we apply a parametric solution that assumes a normal distribution and is extremely fast to train. With a reasonable space and time overhead that is constant in the data size, we set a new state-of-the-art on several public datasets, including a billion-scale one.

Inspired by the close relation between nearest neighbor search and clustering in high-dimensional spaces as well as the success of one helping to solve the other, we introduce a new paradigm where both problems are solved simultaneously. Our solution is recursive, not in the size of input data but in the number of dimensions. One result is a clustering algorithm that is tuned to small codebooks but does not need all data in memory at the same time and is practically constant in the data size. As a by-product, a tree structure performs either exact or approximate quantization on trained centroids, the latter being not very precise but extremely fast. A lesser contribution is a new indexing scheme for image retrieval that exploits multiple small codebooks to provide an arbitrarily fine partition of the descriptor space. Large scale experiments on public datasets exhibit state of the art performance and remarkable generalization.

This paper considers a family of metrics to compare images based on their local descriptors. It encompasses the VLAD descriptor and matching techniques such as Hamming Embedding. Making the bridge between these approaches leads us to propose a match kernel that takes the best of existing techniques by combining an aggregation procedure with a selective match kernel. Finally, the representation underpinning this kernel is approximated, providing a large scale image search both precise and scalable, as shown by our experiments on several benchmarks.

Depending on the application, local feature detectors should comply to properties that are often contradictory, e.g. distinctiveness vs robustness. Providing a good balance is a standing problem in the field. In this direction, we propose a novel approach for local feature detection starting from sampled edges and based on shape stability measures across the weighted alpha-filtration, a computational geometry construction that captures the shape of a non-uniform set of points. Detected features are blob-like and include non-extremal regions as well as regions determined by cavities of boundary shape. The detector provides distinctive regions, while achieving high robustness in terms of repeatability and matching score, as well as competitive performance in a large scale image retrieval application.

We introduce a clustering method that combines the flexibility of Gaussian mixtures with the scaling properties needed to construct visual vocabularies for image retrieval. It is a variant of expectation-maximization that can converge rapidly while dynamically estimating the number of components. We employ approximate nearest neighbor search to speed-up the E-step and exploit its iterative nature to make search incremental, boosting both speed and precision. We achieve superior performance in large scale retrieval, being as fast as the best known approximate k-means.

Many problems, including feature selection, vocabulary learning, location and landmark recognition, structure from motion and 3d reconstruction, rely on a learning process that involves wide-baseline matching on multiple views of the same object or scene. In practical large scale image retrieval applications however, most images depict unique views where this idea does not apply. We exploit self-similarities, symmetries and repeating patterns to select features within a single image. We achieve the same performance compared to the full feature set with only a small fraction of its index size on a dataset of unique views of buildings or urban scenes, in the presence of one million distractors of similar nature. Our best solution is linear in the number of correspondences, with practical running times of just a few milliseconds.

A wide range of properties and assumptions determine the most appropriate spatial matching model for an application, e.g. recognition, detection, registration, or large scale visual search. Most notably, these include discriminative power, geometric invariance, rigidity constraints, mapping constraints, assumptions made on the underlying features or descriptors and, of course, computational complexity. Having image retrieval in mind, we present a very simple model inspired by Hough voting in the transformation space, where votes arise from single feature correspondences. A relaxed matching process allows for multiple matching surfaces or non-rigid objects under one-to-one mapping, yet is linear in the number of correspondences. We apply it to geometry re-ranking in a search engine, yielding superior performance with the same space requirements but a dramatic speed-up compared to the state of the art.

We present a local feature detector that is able to detect regions of arbitrary scale and shape, without scale space construction. We compute a weighted distance map on image gradient, using our exact linear-time algorithm, a variant of group marching for Euclidean space. We find the weighted medial axis by extending residues, typically used in Voronoi skeletons. We decompose the medial axis into a graph representing image structure in terms of peaks and saddle points. A duality property enables reconstruction of regions using the same marching method. We greedily group regions taking both contrast and shape into account. On the way, we select regions according to our shape fragmentation factor, favoring those well enclosed by boundaries--even incomplete. We achieve state of the art performance in matching and retrieval experiments with reduced memory and computational requirements.

We propose a scalable logo recognition approach that extends the common bag-of-words model and incorporates local geometry in the indexing process. Given a query image and a large logo database, the goal is to recognize the logo contained in the query, if any. We locally group features in triples using multi-scale Delaunay triangulation and represent triangles by signatures capturing both visual appearance and local geometry. Each class is represented by the union of such signatures over all instances in the class. We see large scale recognition as a sub-linear search problem where signatures of the query image are looked up in an inverted index structure of the class models. We evaluate our approach on a large-scale logo recognition dataset with more than four thousand classes.

State of the art data mining and image retrieval in community photo collections typically focus on popular subsets, e.g. images containing landmarks or associated to Wikipedia articles. We propose an image clustering scheme that, seen as vector quantization, compresses a large corpus of images by grouping visually consistent ones while providing a guaranteed distortion bound. This allows us, for instance, to represent the visual content of all thousands of images depicting the Parthenon in just a few dozens of scene maps and still be able to retrieve any single, isolated, non-landmark image like a house or a graffiti on a wall. Starting from a geo-tagged dataset, we first group images geographically and then visually, where each visual cluster is assumed to depict different views of the the same scene. We align all views to one reference image and construct a 2D scene map by preserving details from all images while discarding repeating visual features. Our indexing, retrieval and spatial matching scheme then operates directly on scene maps. We evaluate the precision of the proposed method on a challenging one-million urban image dataset.

We present a new approach to image indexing and retrieval, which integrates appearance with global image geometry in the indexing process, while enjoying robustness against viewpoint change, photometric variations, occlusion, and background clutter. We exploit shape parameters of local features to estimate image alignment via a single correspondence. Then, for each feature, we construct a sparse spatial map of all remaining features, encoding their normalized position and appearance, typically vector quantized to visual word. An image is represented by a collection of such feature maps and RANSAC-like matching is reduced to a number of set intersections. Because the induced dissimilarity is still not a metric, we extend min-wise independent permutations to collections of sets and derive a similarity measure for feature map collections. We then exploit sparseness to build an inverted file whereby the retrieval process is sub-linear in the total number of images, ideally linear in the number of relevant ones. We achieve excellent performance on 10^4 images, with a query time in the order of milliseconds.

We believe that the potential of edges in local feature detection has not been fully exploited and therefore propose a detector that starts from single scale edges and produces reliable and interpretable blob-like regions and groups of regions of arbitrary shape. The detector is based on merging local maxima of the distance transform guided by the gradient strength of the surrounding edges. Repeatability and matching score are evaluated and compared to state-of-the-art detectors on standard benchmarks. Furthermore, we demonstrate the potential application of our method to wide-baseline matching and feature detection in sequences involving human activity.

Several spatiotemporal feature point detectors have been recently used in video analysis for action recognition. Feature points are detected using a number of measures, namely saliency, cornerness, periodicity, motion activity etc. Each of these measures is usually intensity-based and provides a different trade-off between density and informativeness. In this paper, we use saliency for feature point detection in videos and incorporate color and motion apart from intensity. Our method uses a multi-scale volumetric representation of the video and involves spatiotemporal operations at the voxel level. Saliency is computed by a global minimization process constrained by pure volumetric constraints, each of them being related to an informative visual aspect, namely spatial proximity, scale and feature similarity (intensity, color, motion). Points are selected as the extrema of the saliency response and prove to balance well between density and informativeness. We provide an intuitive view of the detected points and visual comparisons against state-of-the-art space-time detectors. Our detector outperforms them on the KTH dataset using Nearest-Neighbor classifiers and ranks among the top using different classification frameworks. Statistics and comparisons are also performed on the more difficult Hollywood Human Actions (HOHA) dataset increasing the performance compared to current published results.

Detection of perceptually important video events is formulated here on the basis of saliency models for the audio, visual and textual information conveyed in a video stream. Audio saliency is assessed by cues that quantify multifrequency waveform modulations, extracted through nonlinear operators and energy tracking. Visual saliency is measured through a spatiotemporal attention model driven by intensity, color and motion. Text saliency is extracted from part-of-speech tagging on the subtitles information available with most movie distributions. The various modality curves are integrated in a single attention curve, where the presence of an event may be signified in one or multiple domains. This multimodal saliency curve is the basis of a bottom-up video summarization algorithm, that refines results from unimodal or audiovisual-based skimming. The algorithm performs favorably for video summarization in terms of informativeness and enjoyability.

In this paper we propose a methodology for semantic indexing of images, based on techniques of image segmentation, classification and fuzzy reasoning. The proposed knowledge-assisted analysis architecture integrates algorithms applied on three overlapping levels of semantic information: i) no semantics, i.e. segmentation based on low-level features such as color and shape, ii) mid-level semantics, such as concurrent image segmentation and object detection, region-based classification and, iii) rich semantics, i.e. fuzzy reasoning for extraction of implicit knowledge. In that way, we extract semantic description of raw multimedia content and use it for indexing and retrieval purposes, backed up by a fuzzy knowledge repository. We conducted several experiments to evaluate each technique, as well as the whole methodology in overall and, results show the potential of our approach.

This paper presents an approach on high-level feature detection within video documents, using a Region Thesaurus. A video shot is represented by a single keyframe and MPEG-7 features are extracted locally, from coarse segmented regions. Then a clustering algorithm is applied on those extracted regions and a region thesaurus is constructed to facilitate the description of each keyframe at a higher level than the low-level descriptors but at a lower than the high-level concepts. A model vector representation is formed and several high-level concept detectors are appropriately trained using a global keyframe annotation. The proposed approach is thoroughly evaluated on the TRECVID 2007 development data for the detection of nine high level concepts, demonstrating sufficient performance on large data sets.

A cross media analysis scheme for the semantic interpretation of compound documents is presented. The proposed scheme is essentially a late-fusion mechanism that operates on top of single-media extractors output. Evidence extracted from heterogeneous sources are used to trigger probabilistic inference on a Bayesian network that encodes domain knowledge and quantifies causality. Experiments performed on a set of 54 compound documents showed that the proposed scheme is able to exploit the existing cross media relations and achieve performance improvements.

The popularity of social networks and web-based personal image collections has resulted to a continuously growing volume of publicly available photos and videos. Users are uploading, describing, tagging and annotating their personal photos. Moreover, a recent trend is to also "geotag" them, that is to mark the location they were taken onto a web-based map. Consequently, this growth of image collections has created the need for fast, robust and efficient systems, able to analyze large-scale diverse and heterogeneous visual content. This growing need for automatic metadata generation, concept detection, search and retrieval has boosted research efforts towards these directions. The work presented herein is a web-based system that aims not only to the retrieval of visually similar images, but also to determine the location they were taken by exploiting the available socially created metadata. This system makes use of a visual vocabulary and a bag-of words approach, in order to describe the visual properties of an image. Moreover, geometric constraints are applied, in order to extend the bag-of-words model towards more accurate results. We begin by describing some related work in the field of image retrieval, in order to present both the relation and the novelties of the presented system in comparison with the existing techniques.

Based on perceptual and computational attention modeling studies, we formulate measures of saliency for an audiovisual stream. Audio saliency is captured by signal modulations and related multifrequency band features, extracted through nonlinear operators and energy tracking. Visual saliency is measured by means of a spatiotemporal attention model driven by various feature cues (intensity, color, motion). Audio and video curves are integrated in a single attention curve, where events may be enhanced, suppressed or vanished. The presence of salient events is signified on this audiovisual curve by geometrical features such as local extrema, sharp transition points and level sets. An audiovisual saliency-based movie summarization algorithm is proposed and evaluated. The algorithm is shown to perform very well in terms of summary informativeness and enjoyability for movie clips of various genres.

In this paper we focus on scene classification and detection of high-level concepts within multimedia documents, by introducing an intermediate contextual approach as a means of exploiting the visual context of images. More specifically, we introduce and model a novel relational knowledge representation, founded on topological and semantic relations between the concepts of an image. We further develop an algorithm

to address computationally efficient handling of visual context and extraction of mid-level region characteristics. Based on the proposed knowledge model, we combine the notion of visual context with region semantics, in order to exploit their efficacy in dealing with scene classification problems. Finally, initial experimental results are presented, in order to demonstrate possible applications of the proposed methodology.

In this paper, we propose a framework to extend semantic labeling of images to video shot sequences and achieve efficient and semantic-aware spatiotemporal video segmentation. This task faces two major challenges, namely the temporal variations within a video sequence which affect image segmentation and labeling, and the computational cost of region labeling. Guided by these limitations, we design a method where spatiotemporal segmentation and object labeling are coupled to achieve semantic annotation of video shots. An internal graph structure that describes both visual and semantic properties of image and video regions is adopted. The process of spatiotemporal semantic segmentation is subdivided in two stages: Firstly, the video shot is split into small block of frames. Spatiotemporal regions (volumes) are extracted and labeled individually within each block. Then, we iteratively merge consecutive blocks by a matching procedure which considers both semantic and visual properties. Results on real video sequences show the potential of our approach.

Personalized content retrieval aims at improving the retrieval process by taking into account the particular interests of individual users. However, not all user preferences are relevant in all situations. It is well known that human preferences are complex, multiple, heterogeneous, changing, even contradictory, and should be understood in context with the user goals and tasks at hand. In this paper we propose a method to build a dynamic representation of the semantic context of ongoing retrieval tasks, which is used to activate different subsets of user interests at runtime, in such a way that out of context preferences are discarded. Our approach is based on an ontology-driven representation of the domain of discourse, providing enriched descriptions of the semantics involved in retrieval actions and preferences, and enabling the definition of effective means to relate preferences and context.

In this paper, an affine invariant curve matching method using curvature scale-space and normalization is proposed. Prior to curve matching, curve normalization with respect to affine transformations is applied, allowing a lossless affine invariant curve representation. The maxima points of the curvature scale-space (CSS) image are then used to represent the normalized curve, while retaining the local properties of the curve. The matching algorithm that follows, matches the maxima sets of CSS images and the resulting matching cost provides a measure of similarity. The method's performance and robustness is evaluated through a variety of curves and affine transformations, obtaining precise shape similarity and retrieval.

The notion of context plays a significant role in multimedia content search and retrieval systems. In this paper we focus our research efforts on a visual context knowledge representation, to be utilized for multimedia high-level concept detection. We propose and describe in detail types of contextual relations evident within the multimedia content, model them and provide a clear methodology on how to extract them. A visual context ontology is introduced, containing relations among different types of content entities, such as images, regions, region types and high-level concepts. In this manner, we facilitate traditional object detection approaches towards semantical interpretation. The application of the proposed knowledge structure provides encouraging initial results, improving the efficacy of related multimedia analysis techniques.

This paper presents an approach on high-level feature detection within video documents, using a Region Thesaurus and Latent Semantic Analysis. A video shot is represented by a single keyframe. MPEG-7 features are extracted from coarse regions of it. A clustering algorithm is applied on all extracted regions and a region thesaurus is constructed. Its use is to assist to the mapping of low- to high-level features by a model vector representation. Latent Semantic Analysis is then applied on the model vectors to exploit the latent relations among regions types aiming to improve detection performance. The proposed approach is thoroughly examined using TRECVID 2007 development data.

In this paper we propose the use of enhanced mid-level information, such as information obtained from the application of supervised or unsupervised learning methodologies on low-level characteristics, in order to improve semantic multimedia analysis. High-level, a priori contextual knowledge about the semantic meaning of objects and their low-level visual descriptions are combined in an integrated approach that handles in a uniform way the gap between semantics and low-level features. Prior work on low-level feature extraction is extended and a region thesaurus containing all mid-level features is constructed using a hierarchical clustering method. A model vector that contains the distances from each mid-level element is formed and a Neural Network-based detector is trained for each semantic concept. Contextual adaptation improves the quality of the produced results, by utilizing fuzzy algebra, fuzzy sets and relations. The novelty of the presented work is the context-driven mid-level manipulation of region types, utilizing a domain-independent ontology infrastructure to handle the knowledge. Early experimental results are presented using data derived from the beach domain.

This paper presents an approach for efficient keyframe extraction, using local semantics in form of a region thesaurus. More specifically, certain MPEG-7 color and texture features are locally extracted from keyframe regions. Then, using a hierarchical clustering approach a local region thesaurus is constructed to facilitate the description of each frame in terms of higher semantic features. The feature is consisted by the most common region types that are encountered within the video shot, along with their synonyms. These region types carry semantic information. Each keyframe is represented by a vector consisting of the degrees of confidence of the existence of all region types within this shot. Using this keyframe representation, the most representative keyframe is then selected for each shot. Where a single keyframe is not adequate, using the same algorithm and exploiting the coverage of the visual thesaurus, more keyframes are extracted.

This paper presents an approach on high-level feature detection using a region thesaurus. MPEG-7 features are locally extracted from segmented regions and for a large set of images. A hierarchical clustering approach is applied and a relatively small number of region types is selected. This set of region types defines the region thesaurus. Using this thesaurus, low-level features are mapped to high-level concepts as model vectors. This representation is then used to train support vector machine-based feature detectors. As a next step, latent semantic analysis is applied on the model vectors, to further improve the analysis performance. High-level concepts detected derive from the natural disaster domain.

In this paper we focus on a contextual domain ontology representation aiding in the process of knowledge-assisted multimedia analysis. Previous work on the detection of high-level concepts within multimedia documents is extended by introducing a "mid-level" ontology as a means of exploiting the visual context of images, in terms of high-level concepts and mid-level region types they consist of. More specifically, we introduce a context ontology, define its components, its relations and integrate it in our knowledge modelling approach. In previous works we have developed algorithms to address computationally efficient handling of visual context and extraction of mid-level characteristics and now we expect these diverse algorithms and methodologies to be combined in order to exploit the proposed knowledge model. The ultimate goal remains that of efficient semantic multimedia analysis. Finally, a use case scenario derived from the beach domain is also presented, in order to demonstrate a possible application of the proposed knowledge representation.

The task of multimedia document categorization forms a well-known problem in information retrieval. The task is to assign a multimedia document to one or more categories, based on its contents. In this case, effective management and thematic categorization requires the extraction of the underlying semantics. The proposed approach utilizes as input, analyzes and exploits the textual annotation that accompanies a multimedia document, in order to extract its underlying semantics, construct a semantic index and finally classify the documents to thematic categories. This process is based on a unified knowledge and semantics representation model introduced, as well as basic principles of fuzzy relational algebra. On top of that the fuzzy extension of expressive description logic language SHIN, f-SHIN and its reasoning services are used to further refine and optimize the initial categorization results. The proposed approach was tested on a set of real-life multimedia documents, derived from the Internet, as well as personal databases and shows rather promising results.

A saliency-based method for generating video summaries is presented, which exploits coupled audiovisual information from both media streams. Efficient and advanced speech and image processing algorithms to detect key frames that are acoustically and visually salient are used. Promising results are shown from experiments on a movie database.

In this paper previous work on the detection of high-level concepts within multimedia documents is extended by introducing a mid-level ontology as a means of exploiting the visual context of images in terms of the regions they consist of. More specifically, we construct a mid-level ontology, define its relations and integrate it in our knowledge modelling approach. In the past we have developed algorithms to address computationally efficient handling of visual context and extraction of mid-level characteristics and now we explain how these diverse algorithms and methodologies can be combined in order to approach a greater goal, that of semantic multimedia analysis. Early experimental results are presented using data derived from the beach domain.

This paper describes salienShrink, a method to denoise images based on computing a map of salient coefficients in the wavelet domain and use it to improve common denoising algorithms. By salient, we refer to those coefficients that correspond mostly to pure signal and should therefore be preserved throughout the denoising procedure. We use a computationally efficient model to detect salient regions in the bands of the multiresolution wavelet transform. These regions are used to obtain a more accurate estimate of the noise level, improving the performance of existing well known shrinkage methods. Extensive experimental results on the BiShrink method show that the proposed method effectively enhances PSNR and improves the visual quality of the denoised images.

Event detection and recognition is still one of the most active fields in computer vision, since the complexity of the dynamic events and the need for computational efficient solutions pose several difficulties. This paper addresses detection and representation of spatiotemporal salient regions using the 3D Discrete Wavelet Transform (DWT). We propose a framework to measure saliency based on the orientation selective bands of the 3D DWT and represent events using simple features of salient regions. We apply this method to human action recognition, test it on a large public video database consisting of six human actions and compare the results against an established method in the literature. Qualitative and quantitative evaluation indicates the potential of the proposed method to localize and represent human actions.

In this paper we propose an algorithm to improve the results of knowledge-assisted image analysis, based on contextual information. In order to achieve this, we utilize fuzzy algebra, fuzzy sets and relations, towards efficient manipulation of image region concepts. We provide a novel context modelling, based on the OWL language, using RDF reification. Initial image analysis results are enhanced by the utilization of domain-independent, semantic knowledge in terms of concepts and relations between them. The novelty of the presented work is the context-driven re-adjustment of the degrees of confidence of the detected concepts produced by any image analysis technique, utilizing a domain-independent ontology infrastructure to handle the knowledge, as well as multiple application domains.

In this paper, we propose the application of rule-based reasoning for knowledge assisted image segmentation and object detection. A region merging approach is proposed based on fuzzy labeling and not on visual descriptors, while reasoning is used in evaluation of dissimilarity between adjacent regions according to rules applied on local information.

In this paper we present a framework for simultaneous image segmentation and region labeling leading to automatic image annotation. The proposed framework operates at semantic level using possible semantic labels to make decisions on handling image regions instead of visual features used traditionally. In order to stress its independence of a specific image segmentation approach we applied our idea on two region growing algorithms, i.e. watershed and recursive shortest spanning tree. Additionally we exploit the notion of visual context by employing fuzzy algebra and ontological taxonomic knowledge representation, incorporating in this way global information and improving region interpretation. In this process, semantic region growing labeling results are being re-adjusted appropriately, utilizing contextual knowledge in the form of domain-specific semantic concepts and relations. The performance of the overall methodology is demonstrated on a real-life still image dataset from the popular domains of beach holidays and motorsports.

In this poster, we present an approach to contextualized semantic image annotation as an optimization problem. Ontologies are used to capture general and contextual knowledge of the domain considered, and a genetic algorithm is applied to realize the final annotation. Experiments with images from the beach vacation domain demonstrate the performance of the proposed approach and illustrate the added value of utilizing contextual information.

This paper presents a framework for the detection of semantic features in video sequences. Low-level feature extraction is performed on the keyframes of the shots and a "feature vector" including color and texture features is formed. A region "thesaurus" that contains all the high-level features is constructed using a subtractive clustering method.Then, a "model vector" that contains the distances from each region type is formed and a SVM detector is trained for each semantic concept. Experiments were performed using TRECVID 2005 development data.

Tackling the problems of automatic object recognition and/or scene classification with generic algorithms is not producing efficient and reliable results in the field of image analysis. Restricting the problem to a specific domain is a common approach to cope with this, still unresolved, issue. In this paper we propose a methodology to improve the results of image analysis, based on available contextual information derived from the popular sports domain. Our research efforts include application of a knowledge-assisted image analysis algorithm that utilizes an ontology infrastructure to handle knowledge and MPEG-7 visual descriptors for region labeling. A novel ontological representation for context is introduced, combining fuzziness with Semantic Web characteristics, such as RDF. Initial region labeling analysis results are then being re-adjusted appropriately according to a confidence value readjustment algorithm, by means of fine-tuning the degrees of confidence of each detected region label. In this process contextual knowledge in the form of domain-specific semantic concepts and relations is utilized. Performance of the overall methodology is demonstrated through its application on a real-life still image dataset derived from the tennis sub-domain.

In this paper we investigate the utilization of visual saliency maps for ROI-based video coding of video-telephony applications. Visually salient areas indicated in the saliency map are considered as ROIs. These areas are automatically detected using an algorithm for visual attention (VA) which builds on the bottom-up approach proposed by Itti et al. A top-down channel emulating the visual search for human faces performed by humans has been added, while orientation, intensity and color conspicuity maps are computed within a unified multi-resolution framework based on wavelet subband analysis. Priority encoding, for experimentation purposes, is utilized in a simple manner: Frame areas outside the priority regions are blurred using a smoothing filter and then passed to the video encoder. This leads to better compression of both Intra-coded (I) frames (more DCT coefficients are zeroed in the DCT-quantization step) and Inter coded (P,B) frames (lower prediction error). In more sophisticated approaches, priority encoding could be incorporated by varying the quality factor of the DCT quantization table. Extended experiments concerning both static images as well as low-quality video show the compression efficiency of the proposed method. The comparisons are made against standard JPEG and MPEG-1 encoding respectively.

Combining traditional personalization techniques with novel knowledge representation paradigms, such as the ontology-based approach proposed in the Semantic Web field, is a challenging task. Personalization is a difficult problem when dealing with multimedia content and information retrieval, where context is increasingly acknowledged to be a key notion in order to make proper sense of user needs. This work focuses on contextualization within personalization in a multimedia environment. Towards that scope, we propose a novel contextual knowledge modeling scheme, and an approach for the dynamic, contextual activation of semantic user preferences to better represent user interests in coherence with ongoing user activities, e.g. in an interactive retrieval process. The application of this methodology is demonstrated using two user scenarios, and the performance results of a preliminary experiment are shown.

Personalized content retrieval aims at improving the retrieval process by taking into account the particular interests of individual users. However, not all user preferences are relevant in all situations. It is well known that human preferences are complex, multiple, heterogeneous, changing, even contradictory, and should be understood in context with the user goals and tasks at hand. In this paper we propose a method to build a dynamic representation of the semantic context of ongoing retrieval tasks, which is used to activate different subsets of user interests at runtime, in such a way that out of context preferences are discarded. Our approach is based on an ontology-driven representation of the domain of discourse, providing enriched descriptions of the semantics involved in retrieval actions and preferences, and enabling the definition of effective means to relate preferences and context.

Personalization is a difficult problem related to fields and applications ranging from information retrieval to multimedia content manipulation. Challenge is greater, when trying to combine traditional personalization techniques with novel knowledge representations like ontologies. This paper proposes a novel contextual knowledge modeling, based on ontologies and fuzzy relations and exploits it in user profiling representation, extraction and use. The personalized results of the application of this methodology are then ranked accordingly. The performance of the proposed techniques is demonstrated through preliminary experimental results derived from a real-life data set.

One of the most common problems in computer vision and image processing applications is the localization of object boundaries in a video frame and its tracking in the next frames. In this paper, a fully automatic method for fast tracking of video objects in a video sequence using affine invariant normalization is proposed. Initially, the detection of a video object is achieved using a GVF snake. Next, a vector of the affine parameters of each contour of the extracted video object in two successive frames is computed using affine-invariant normalization. Under the hypothesis that these contours are similar, the affine transformation between the two contours is computed in a very fast way. Using this transformation to predict the position of the contour in the next frame allows initialization of the GVF snake very close to the real position. Applying this technique to the following frames, a very fast tracking technique is achieved. Moreover, this technique can be applied on sequences with very fast moving objects where traditional trackers usually fail. Results on synthetic sequences are presented which illustrate the theoretical developments.

In this position paper we examine the limitation of region growing segmentation techniques to extract semantically meaningful objects from an image. We propose a region growing algorithm that performs on a semantic level, driven by the knowledge of what each region represents at every iteration step of the merging process. This approach utilizes simultaneous segmentation and labeling of regions leading to automatic image annotation.

Generic algorithms for automatic object recognition and/or scene classification are unfortunately not producing reliable and robust results. A common approach to cope with this, still unresolved, issue is to restrict the problem at hand to a specific domain. In this paper we propose an algorithm to improve the results of image analysis, based on the contextual information we have, which relates the detected concepts to any given domain. Initial results produced by the image analysis module are domain-specific semantic concepts and are being re-adjusted appropriately by the suggested algorithm, by means of fine-tuning the degrees of confidence of each detected concept. The novelty of the presented work is twofold: i) the knowledge-assisted image analysis algorithm, that utilizes an ontology infrastructure to handle the knowledge and MPEG-7 visual descriptors for the region labeling and ii) the context-driven re-adjustment of the degrees of confidence of the detected labels.

This paper proposes a new type of a support vector machine which uses a kernel constituted from fuzzy basis functions. The proposed network combines the characteristics both of a support vector machine and a fuzzy system: high generalization performance, even when the dimension of the input space is very high, structured and numerical representation of knowledge and ability to extract linguistic fuzzy rules, in order to bridge the "semantic gap" between the low-level descriptors and the high-level semantics of an image. The Fuzzy SVM network was evaluated using images from the aceMedia Repository and more specifically in a beach/urban scenes classification problem.

In this paper we discuss the use of knowledge for the automatic extraction of semantic metadata from multimedia content. For the representation of knowledge we extended and enriched current general-purpose ontologies to include low-level visual features. More specifically, we implemented a tool that links MPEG-7 visual descriptors to high-level, domain-specific concepts. For the exploitation of this knowledge infrastructure we developed an experimentation platform, that allows us to analyze multimedia content and automatically create the associated semantic metadata, as well as to test, validate and refine the ontologies built. We pursued a tight and functional integration of the knowledge base and the analysis modules putting them in a loop of constant interaction instead of being the one just a pre- or post-processing step of the other.

Knowledge representation and annotation of multimedia documents typically have been pursued in two different directions. Previous approaches have focused either on low level descriptors, such as dominant color, or on the content dimension and corresponding manual annotations, such as person or vehicle. In this paper, we present a knowledge infrastructure to bridge the gap between the two directions. Ontologies are being extended and enriched to include low-level audiovisual features and descriptors. Additionally, a tool for linking low-level MPEG-7 visual descriptions to ontologies and annotations has been developed. In this way, we construct ontologies that include prototypical instances of domain concepts together with a formal specification of the corresponding visual descriptors. Thus, we combine high-level domain concepts and low-level multimedia descriptions, enabling for new media content analysis.

Reliability is a well-known concern in the field of personalization technologies. We propose the extension of an ontology-based retrieval system with semantic-based personalization techniques, upon which automatic mechanisms are devised that dynamically gauge the degree of personalization, so as to benefit from adaptivity but yet reduce the risk of obtrusiveness and loss of user control. On the basis of a common domain ontology KB, the personalization framework represents, captures and exploits user preferences to bias search results towards personal user interests. Upon this, the intensity of personalization is automatically increased or decreased according to an assessment of the imprecision contained in user requests and system responses before personalization is applied.

It is common sense among experts that visual attention plays an important role in perception, being necessary for obtaining salient information about the surroundings. It may be the "glue" that binds simple visual features into an object [1]. Having proposed a spatiotemporal model for visual attention in the past, we elaborate on this work and use it for video classification. Our claim is that simple visual features bound to spatiotemporal salient regions will better represent the video content. Hence, we expect that feature vectors extracted from these regions will enhance the performance of the classifier. We present statistics on sports sequences of five different categories that verify our claims.

In this paper, an ontology infrastucture for multimedia reasoning is presented, making it possible to combine low-level visual descriptors with domain specific knowledge and subsequently analyze multimedia content with a generic algorithm that makes use of this knowledge. More specifically, the ontology infrastructure consists of a domain-specific ontology, a visual descriptor ontology (VDO) and an upper ontology. In order to interpret a scene, a set of atom regions is generated by an initial segmentation and their descriptors are extracted. Considering all descriptors in association with the related prototype instances and relations, a genetic algorithm labels the atom regions. Finally, a constraint reasoning engine enables the final region merging and labelling into meaningful objects.

This paper describes a comprehensive framework giving support to a wide range of personalization facilities in a multi-media content management environment. The framework builds upon a rich, ontology-based representation of the domain of discourse, whereby content semantics are linked to a rich representation of user preferences. The expressive power of ontologies is used to develop automatic learning capabilities, in order to update user profiles as users interact with the system. The resulting descriptions of user interests in terms of ontologies are exploited, along with available content metadata, to provide users with personalized content search, browsing, ranking, and retrieval. On a wider perspective, the framework is built as an open platform that provides for further user and device adaptive capability extensions.

This paper proposes a number of content-based image classification techniques based on fusing various low-level MPEG-7 visual descriptors. The goal is to fuse several descriptors in order to improve the performance of several machine-learning classifiers. Fusion is necessary as descriptors would be otherwise incompatible and inappropriate to directly include e.g. in a Euclidean distance. Three approaches are described: A merging fusion combined with an SVM classifier, a back-propagation fusion combined with a K-Nearest Neighbor classifier and a Fuzzy-ART neurofuzzy network. In the latter case, fuzzy rules can be extracted in an effort to bridge the semantic gap between the low-level descriptors and the high-level semantics of an image. All networks were evaluated using content from the aceMedia Repository and more specifically in a beach/urban scenes classification problem.

Context is of great importance in a wide range of computing applications and has become a major topic in multimedia content search and retrieval systems. In this paper we focus our research efforts on visual context, a part of context suitable for multimedia analysis and usage. We introduce our efforts towards the scope of clarifying context in the fields of object detection and scene classification during multimedia analysis. We also present a method for visual context modelling, based on spatial object and region-based relations, to use in content-based multimedia search and retrieval systems.

Inspired by the human visual system, visual attention (VA) models seem to provide solutions to problems of semantic image understanding by selecting only a small but representative fraction of visual input to process. Having proposed a spatiotemporal VA model for video processing in the past, we propose considerable enhancements in this paper, including the use of steerable filters for 3D orientation estimation, and of PCA for fusion of features for the construction of saliency volumes. We further employ segmentation and feature extraction on salient regions to provide video classification using an SVM classifier. Finally, we provide results on sports video classification and comment on the usefulness of spatiotemporal VA for such purposes.

Annotations of multimedia documents typically have been pursued in two different directions. Either previous approaches have focused on low level descriptors, such as dominant color, or they have focused on the content dimension and corresponding annotations, such as person or vehicle. In this paper, we present a software environment to bridge between the two directions. M-OntoMat-Annotizer allows for linking low level MPEG-7 visual descriptions to conventional Semantic Web ontologies and annotations. We use M-OntoMat-Annotizer in order to construct ontologies that include prototypical instances of high-level domain concepts together with a formal specification of corresponding visual descriptors. Thus, we formalize the interrelationship of high- and low-level multimedia concept descriptions allowing for new kinds of multimedia content analysis and reasoning.

In this paper we present the construction of an ontology that represents the structure of the MPEG-7 visual part. The goal of this ontology is to enable machines to generate and understand visual descriptions which can be used for multimedia reasoning. Within the specification, MPEG-7 definitions (description schemes and descriptors) are expressed in XML Schema. Although XML Schema provides the syntactic, structural, cardinality and datatyping constraints required by MPEG-7, it does not provide the semantic interoperability required to make MPEG-7 visual descriptors accessible by other domains. The knowledge representation provided by the ontology can be used to develop tools which perform knowledge-based reasoning. For the construction of the ontology we use the RDFS ontology language. We present the problems that occurred, mainly, due to the RDFS modelling limitations. Finally, we propose a way to apply reasoning using the VD ontology.

Efficient video content management and exploitation requires extraction of the underlying semantics, which is a non-trivial task involving the association of low-level features with high-level concepts. In this paper, a knowledge-assisted approach for extracting semantic information of domain-specific video content is presented. Domain knowledge considers both low-level visual features (color, motion, shape) and spatial information (topological and directional relations). An initial segmentation algorithm generates a set of over-segmented atom-regions and a neural network is used to estimate the similarity distance between the extracted atom-region descriptors and the ones of the object models included in the domain ontology. A genetic algorithm is applied then in order to find the optimal interpretation according to the domain conceptualization. The proposed approach was tested on the Tennis and Formula One domains with promising results.

In natural vision, we center our fixation on the most informative points in a scene in order to reduce our overall uncertainty about the scene and help interpret it. Even if we are looking for a specific stimulus around us, we face a great amount of uncertainty since that stimulus could be in any spatial location. Visual attention (VA) schemes have been proposed by researchers to account for the ability of the human eye to quickly fixate on informative regions. Recently, VA in images, and especially saliency-based VA, became an active research topic of the computer vision community. The proposed work provides an extension towards VA in video sequences by integrating spatiotemporal information. The potential applications include video classification, scene understanding, surveillance and segmentation.

In this paper, an integrated information system is presented that offers enhanced search and retrieval capabilities to users of hetero-lingual digital audiovisual (a/v) archives. This innovative system exploits the advances in handling a/v content and related metadata, as introduced by MPEG-4 and worked out by MPEG-7, to offer advanced services characterized by the tri-fold "semantic phrasing of the request (query)", "unified handling" and "personalized response". The proposed system is targeting the intelligent extraction of semantic information from a/v and text related data taking into account the nature of the queries that users my issue, and the context determined by user profiles.

Integration of knowledge and multimedia content technologies is important for the future of European industry and commerce. aceMedia is an IST FP6 project which aims to unite these two established disciplines to achieve significant advances by the combination of the two domains. This paper describes research in content processing and knowledge assisted multimedia analysis within the aceMedia project, and provides a scenario of use which illustrates the benefits of this combined approach.

In this paper, a knowledge representation infrastructure for semantic multimedia content analysis and reasoning is presented. This is one of the major objectives of the aceMedia Integrated Project where ontologies are being extended and enriched to include low-level audiovisual features, descriptors and behavioural models in order to support automatic content annotation. More specifically, the developed infrastructure consists of the core ontology based on extensions of the DOLCE core ontology and the multimedia-specific infrastructure components. These are, the Visual Descriptors Ontology, which is based on an RDFS representation of the MPEG-7 Visual Descriptors and the Multimedia Structure Ontology, based on the MPEG-7 MDS. Furthermore, the developed Visual Descriptor Extraction tool is presented, which will support the initialization of domain ontologies with multimedia features.

Several visual attention (VA) schemes have been proposed with the saliency-based ones being the most popular. The proposed work provides an extension towards VA video sequences by treating it as volumetric data. The architecture is presented in detail and potential applications are investigated. We expect that the extended VA scheme will reveal interesting events across the sequence like occlusions and short occurrences of objects, providing a basis for video surveillance (e.g. intruder detection) and summarization applications.

In this paper we follow a fuzzy relational approach to knowledge representation. With the use of semantic fuzzy relations we define and extract the semantic context out of a set of semantic entities. Based on this, we then proceed to the case of information retrieval and explain how the three participating contexts, namely the context of the query, the context of the document and the context of the user, can be estimated and utilized towards the achievement of more intuitive information services.

This paper presents FAETHON, a distributed information system that offers enhanced search and retrieval capabilities to users interacting with digital audiovisual (a/v) archives. Its novelty primarily originates in the unified intelligent access to heterogeneous a/v content. The paper emphasizes on the features that provide enhanced search and retrieval capabilities to users, as well as intelligent management of the a/v content by content creators / distributors. It describes the system's main components, the intelligent metadata creation package, the a/v search engine & portal, and the MPEG-7 compliant a/v archive interfaces. Finally, it provides ideas on the positioning of FAETHON in the market of a/v archives and video indexing and retrieval.

In this paper we discuss the use of knowledge for the analysis and semantic retrieval of video. We follow a fuzzy relational approach to knowledge representation, based on which we define and extract the context of either a multimedia document or a user query. During indexing, the context of the document is utilized for the detection of objects and for automatic thematic categorization. During retrieval, the context of the query is used to clarify the exact meaning of the query terms and to meaningfully guide the process of query expansion and index matching. Indexing and retrieval tools have been implemented to demonstrate the proposed techniques and results are presented using video from audiovisual archives.

In this paper, an approach for knowledge and context-assisted content analysis and reasoning based on a multimedia ontology infrastructure is presented. This is one of the major objectives of the aceMedia Integrated Project. In aceMedia, ontologies will be extended and enriched to include lowlevel audiovisual features, descriptors and behavioural models in order to support automatic content annotation. This approach is part of an integrated framework consisting of: user-oriented design, knowledge-driven content processing and distributed system architecture. The overall objective of aceMedia is the implementation of a novel concept for unified media representation: the Autonomous Content Entity (ACE), which has three layers: content, its associated metadata, and an intelligence layer. The ACE concept will be verified by two user focused application prototypes, enabled for both home network and mobile communication environments.

Extraction of visual descriptor is a crucial problem for state-of-the-art visual information analysis. In this paper, we present a knowledge-based approach for detection of visual objects in video sequences. The propose approach models objects through their visual descriptors defined in MPEG7. It first extracts moving regions using an efficient active contours technique. It then computes visual descriptions of the moving regions including color features, shape features which are invariant to affine transformations, as well as motion features. The extracted features are matched to a-priori knowledge about the objects' descriptions,using appropriately defined matching functions. Results are presented which illustrate the theoretical developments

Object detection techniques are coming closer to the automatic detection and identification of objects in multimedia documents. Still, this is not sufficient for the understanding of multimedia content, mainly because a simple object may be related to multiple topics, few of which are indeed related to a given document. In this paper we determine the thematic categories that are related to a document based on the objects that have been automatically detected in it. Our approach relies on stored knowledge and a fuzzy hierarchical clustering algorithm; this algorithm uses a similarity measure that is based on the notion of context. The context is extracted using fuzzy ontological relations.

In this paper, an integrated information system is presented that offers enhanced search and retrieval capabilities to users of heterogeneous digital audiovisual (a/v) archives. This innovative system exploits the advances in handlings a/v content and related metadata, as introduced by MPEG-4 and worked out by MPEG-7, to offer advanced services characterized by the tri-fold "semantic phrasing of the request (query)", "unified handling" and "personalized response". The proposed system is targeting the intelligent extraction of semantic information from a/v and text related data taking into account the nature of the queries that users my issue, and the context determined by user profiles. It also provides a personalisation process of the response in order to provide end-users with desired information. From a technical point of view, the FAETHON system plays the role of an intermediate access server residing between the end users and multiple heterogeneous audiovisual archives organized according to the new MPEG standards.

Multimedia Content is described via textual, semantic and structural Descriptors and Description Schemes, as introduced in MPEG-7. The semantic part of the description is closer to what the user expects from a multimedia search engine, however it poses difficulties, because of the potential incompatibility of the semantic entities among different archives. In this paper, we present FAETHON, a system that unifies the semantic description of heterogeneous archives, through the use of a semantic encyclopaedia.

In this paper, an integrated information system is presented that offers enhanced search and retrieval capabilities to users of heterogeneous digital audiovisual (a/v) archives. This novel system exploits the advances in handling a/v content and related metadata, as introduced by MPEG-4 and worked out by MPEG-7, to offer advanced access services characterized by the tri-fold "semantic phrasing of the request (query)", "unified handling" and "personalized response". The proposed system is targeting the intelligent extraction of semantic information from a/v and text related data taking into account the nature of useful queries that users may issue, and the context determined by user profiles. From a technical point of view, it will play the role of an intermediate access server residing between the end users and multiple heterogeneous audiovisual archives organized according to new MPEG standards.

A system for digitization, storage and retrieval of audiovisual information and its associated data (meta-info) is presented. The principles of the evolving MPEG-7 standard have been adopted for the creation of the data model used by the system, permitting efficient separation of database design, content description, business logic and presentation of query results. XML Schema is used in defining the data model, and XML in describing audiovisual content. Issues regarding problems that emerged during system design and their solutions are discussed, such as customization, deviations from the standard MPEG-7 DSs or even the design of entirely custom DSs. Although the system includes modules for digitization, annotation, archiving and intelligent data mining, the paper mainly focuses on the use of MPEG-7 as the information model.

In this paper, the FAETHON project is presented, whose overall objective is to develop an integrated information system that offers enhanced search and retrieval capabilities to users of heterogeneous digital audiovisual (a/v) archives. This novel system will exploit the advances in handling a/v content and related metadata, as introduced by MPEG-4 and worked out by MPEG-7, to offer advanced access services characterized by the tri-fold "semantic phrasing of the request (query)", "unified handling" and "personalized response".

Content-based audiovisual data retrieval utilizing new emerging related standards such as MPEG-7 will yield ineffective results, unless major focus is given to the semantic information level. Mapping of low level, sub-symbolic descriptors of a/v archives to high level symbolic ones is in general difficult, even impossible with the current state of technology. It can, however, be tackled when dealing with specific application domains. It seems that the extraction of semantic information from a/v and text related data is tractable taking into account the nature of useful queries that users may issue and the context determined by user profile. The IST project FAETHON is developing a novel platform to implement these ideas for user friendly and highly informative access to distributed audiovisual archives.

A novel method for two-dimensional curve normalization with respect to affine transformations is presented in this paper, allowing an affine-invariant curve representation to be obtained without any actual loss of information on the original curve. It can be applied as a pre-processing step to any shape representation, classification, recognition or retrieval technique, since it effectively decouples the problem of affine-invariant description from feature extraction and pattern matching. Curves estimated from object contours are first modeled by cubic B-splines and then normalized in several steps in order to eliminate translation, scaling, skew, starting point, rotation and reflection transformations, based on a combination of curve features including moments and Fourier descriptors.

Face detection is becoming an important tool in the framework of many multimedia applications. Several face detection algorithms based on skin color characteristics have recently appeared in the literature. Most of them face generalization problems due to the skin color model they use. In this work we present a study which attempts to minimize the generalization problem by combining the M-RSST color segmentation algorithm with a Gaussian model of the skin color distribution and global shape features. Moreover by associating the resultant segments with a face probability we can index and retrieve facial images from multimedia databases.

Content-based retrieval from image databases attracts increasing interest the last few years. On the other hand several recent works on face detection based on the chrominance components of the color space have been presented in the literature showing promising results. In this work we combine color segmentation techniques and color based face detection in an efficient way for the purpose of facial image retrieving. In particular, images stored in a multimedia database are analyzed using the M-RSST segmentation algorithm and segment features including average color components, size, location, shape and texture are extracted for several image resolutions. An adaptive two-dimensional Gaussian density function is then employed for modeling skin-tone chrominance color component distribution and detecting image segments that probably correspond to human faces. This information is combined with object shape characteristics so that robust face detection is achieved. Based on the above, a query by example framework is proposed, supporting a highly interactive, configurable and flexible content-based retrieval system for human faces. Experimental results have shown that the proposed implementation combines efficiency, robustness and speed, and could be extended to generic visual information retrieval or video databases.

Automatic content-based analysis and indexing of broadcast news recordings or digitized news archives is becoming an important tool in the framework of many multimedia interactive services such as news summarization, browsing, retrieval and news-on-demand (NoD) applications. Existing approaches have achieved high performance in such applications but heavily rely on textual cues such as closed caption tokens and teletext transcripts. In this work we present an efficient technique for temporal segmentation and parsing of news recordings based on visual cues that can either be employed as stand-alone application for non-closed captioned broadcasts or integrated with audio and textual cues of existing systems. The technique involves robust face detection by means of color segmentation, skin color matching and shape processing, and is able to identify typical news instances like anchorpersons, reports and outdoor shots.

A video content representation framework is proposed in this paper for extracting limited but meaningful information of video data directly from MPEG compressed domain. A hierarchical color and motion segmentation scheme is applied to each video shot, transforming the conventional frame-based representation to a feature-based one. Then, all features are gathered together using a fuzzy formulation and extraction of several key frames is performed for each shot in a content-based rate sampling framework. In particular, our approach is based on minimization of a cross-correlation criterion among video frames of a given shot, so as to be located a set of minimally correlated feature vectors.

In the context of this paper a generalized framework for non-linear representation of 3-D video sequences is proposed, regardless of the scene complexity. In particular, depth information is exploited to provide a more reliable video content segmentation. In this paper this is accomplished by merging color segments which belong to similar depth since a video object is usually located on the same depth plane while color segments give very accurate contours of the objects. To accelerate the segmentation algorithm a multiresolution implementation of the Recursive Shortest Spanning Tree (RSST) algorithm is presented both for color and depth segmentation. All features extracted by the video sequence analysis module are gathered together using a fuzzy feature vector formulation to increase the robustness of the proposed summarization scheme. Finally, key frames within each shot are extracted by minimizing a cross correlation criterion by means of a genetic algorithm.

A facial expression recognition scheme is presented in this paper, based on features derived from the optical flow between two instances of a face in the same emotional state. A pre-processing step of isolating the human face from the background is first employed by means of face detection and registration. A spatio-temporal description of the expression is then obtained by evaluating the Radon transform of the motion vectors between the face in its neutral condition and at the 'apex' of the expression. A linear curve normalization scheme is proposed, achieving a translation, scaling and resolution invariant representation of the Radon curves. Finally, experimental results are presented, illustrating the performance of the proposed algorithm for expression classification using a correlation criterion and a neural network classifier.

Recent literature comprises a large number of papers on the query and retrieval of visual information based on its content. At the same time, a number of prototype systems have been implemented enabling searching through on-line image databases and still image retrieval. However, it has been often pointed out that meaningful/semantic information should be extracted from visual information in order to improve the efficiency and functionality of a content-based retrieval tool. In this context, present work focuses on the extraction of objects from images and video clips and modeling of the resulting object contours using B-splines. Affine-invariant curve representation is obtained through Normalized Fourier descriptors (NFD), curve moments, as well as a novel curve normalization algorithm that leads to major preservation of object shape information. A neural network approach is employed for supervised classification of video objects into prototype object classes. Experiments on several real-life and simulated video sequences are included to evaluate the classification results for all affine-invariant representations used.

This paper presents an integrated framework for interactive content-­based retrieval in video databases by means of visual queries. The proposed system incorporates algorithms for video shot detection, key­ frame and shot selection, automated video object segmentation and tracking, and construction of multidimensional feature vectors using fuzzy classification of color, motion or texture segment properties. Retrieval is then performed in an interactive way by employing a parametric distance between feature vectors and updating distance parameters according to user requirements using relevance feedback. Experimental results demonstrate increased performance and flexibility according to user information needs.

In this paper, a system for content-based image retrieval from video databases is introduced, using B-splines for affine invariant object representation. A small number of key-frames is extracted from each video sequence, which provide sufficient information about the video content. Color and motion segmentation and tracking is then employed for automatic extraction of video objects. A B-spline representation of the object contours is then obtained, which possesses important properties, such as smoothness, continuity and invariance under affine transformation. A neural network approach is used for supervised classification of video objects into prototype object classes. Finally, higher level classes can be constructed combining primary classes, providing the ability to obtain a high level of abstraction in the representation of each video sequence.

An integrated framework for content-based indexing and retrieval in video databases is presented in this paper, which has the capability of adapting its performance according to user requirements. Video sequences are represented by extracting a small number of key frames or scenes and constructing multidimensional feature vectors using fuzzy classification of color, motion or texture segment properties. Queries are then performed by employing a parametric distance between feature vectors, and adaptation is achieved by estimating distance parameters according to user requirements, resulting in a content based retrieval system of increased performance and flexibility.

In this paper, an efficient video content representation is proposed using optimal extraction of characteristic frames and scenes. This representation, apart from providing browsing capabilities to digital video databases, also allows more efficient content-based queries and indexing. For performing the frame/scene extraction, a feature vector formulation of the images is proposed based on color and motion segmentation. Then, the scene selection is accomplished by clustering similar scenes based on a distortion criterion. Frame selection is performed using an optimization method for locating a set of minimally correlated feature vectors.

A new method for designing ultrasonic imaging systems is presented in this paper. The method is based on the use of transducer arrays whose elements transmit wideband signals generated by pseudo-random codes, similarly to code division multiple access (CDMA) systems in communications. The use of code sequences instead of pulses, which are typically used in conventional phased arrays, combined with transmit and receive beamforming for steering different codes at each direction, permits parallel acquisition of a large number of measurements corresponding to different directions. Significantly higher image acquisition rate as well as lateral and contrast resolution are thus obtained, while axial resolution remains close to that of phased arrays operating in pulse-echo mode. Time and frequency division techniques are also studied and a unified theoretical model is derived, which is validated by experimental results.

The first stage of the proposed algorithm includes a scene cut detection mechanism. Then, video processing and image analysis techniques are applied to each video frame for extracting color, motion and texture information. Color information is extracted by applying a hierarchical color segmentation algorithm to each video frame. Consequently, apart from the color histogram of each frame additional features are collected concerning the number of color segments, and their location, size and shape. Motion information is also extracted in a similar way by using a motion estimation and segmentation algorithm. All the above features are gathered in order to form a multidimensional feature vector for each video frame. The representation of each frame by a feature vector, apart from reducing storage requirements, transforms the image domain to another domain, more efficient for key frame selection. Since similar frames can be characterized by different color or motion segments, due to imperfections of the segmentation algorithms, a fuzzy representation of feature vectors is adopted in order to provide more robust searching capabilities. In particular, we classify color as well as motion and texture segments into pre-determined classes forming a multidimensional histogram and a degree of membership is allocated to each category so that the possibility of erroneous comparisons is eliminated.

In this paper, an efficient video content representation system is presented which permits automatic extraction of a limited number of characteristic frames or scenes that provide sufficient information about the content of an MPEG video sequence. This can be used for reduction of the amount of stored information that is necessary in order to provide search capabilities in a multimedia database, resulting in faster and more efficient video queries. Moreover, the proposed system can be used for automatic generation of low resolution video clip previews (trailers), giving the ability to browse databases on web pages. Finally, direct content-­based retrieval with image queries is possible using the feature vector representation incorporated in our system.

Recent progress in supervised image classification research, has demonstrated the potential usefulness of incorporating fuzziness in the training, allocation and testing stages of several classification techniques. In this paper a multiresolution neural network approach to supervised classification is presented, exploiting the inherent fuzziness of such techniques in order to perform classification at different resolution levels and gain in computational complexity. In particular, multiresolution image analysis is carried out and hierarchical neural networks are used as an efficient architecture for classification of the derived multiresolution image representations. A new scheme is then introduced for transferring classification results to higher resolutions based on the fuzziness of the results of lower resolutions, resulting in faster implementation. Experimental results on land cover mapping applications from remotely sensed data illustrate significant improvements in classification speed without deterioration of representation accuracy.

An integrated framework for automatic extraction of the most characteristic frames or scenes of a video sequence is presented in this paper. This is accomplished by extracting a collection of a small number of frames or scenes that provide sufficient information about the video sequence. The scene/frame selection mechanism is based on a transformation from the image to a feature domain, which is more suitable for image comparisons, queries and retrieval.

The implementation of an efficient scheme for translation, rotation and scale invariant optical character recognition is presented in this paper. An image representation is used, which is based on appropriate clustering and transformation of the image triple-correlation domain. This representation is one-to-one related to the class of all shifted-rotated-scaled versions of the original image, as well as robust to a wide variety of additive noises. Special attention is given to binary images, which are used for Optical Character Recognition, and simulation results illustrate the performance of the proposed implementation.

This work introduces composed image retrieval to remote sensing. It allows to query a large image archive by image examples alternated by a textual description, enriching the descriptive power over unimodal queries, either visual or textual. Various attributes can be modified by the textual part, such as shape, color, or context. A novel method fusing image-to-image and text-to-image similarity is introduced. We demonstrate that a vision-language model possesses sufficient descriptive power and no further learning step or training data are necessary. We present a new evaluation benchmark focused on color, context, density, existence, quantity, and shape modifications. Our work not only sets the state-of-the-art for this task, but also serves as a foundational step in addressing a gap in the field of remote sensing image retrieval. Code at: this https URL.

Multi-target unsupervised domain adaptation (UDA) aims to learn a unified model to address the domain shift between multiple target domains. Due to the difficulty of obtaining annotations for dense predictions, it has recently been introduced into cross-domain semantic segmentation. However, most existing solutions require labeled data from the source domain and unlabeled data from multiple target domains concurrently during training. Collectively, we refer to this data as "external". When faced with new unlabeled data from an unseen target domain, these solutions either do not generalize well or require retraining from scratch on all data. To address these challenges, we introduce a new strategy called "multi-target UDA without external data" for semantic segmentation. Specifically, the segmentation model is initially trained on the external data. Then, it is adapted to a new unseen target domain without accessing any external data. This approach is thus more scalable than existing solutions and remains applicable when external data is inaccessible. We demonstrate this strategy using a simple method that incorporates self-distillation and adversarial learning, where knowledge acquired from the external data is preserved during adaptation through "one-way" adversarial learning. Extensive experiments in several synthetic-to-real and real-to-real adaptation settings on four benchmark urban driving datasets show that our method significantly outperforms current state-of-the-art solutions, even in the absence of external data. Our source code is available online this https URL.

This paper studies interpretability of convolutional networks by means of saliency maps. Most approaches based on Class Activation Maps (CAM) combine information from fully connected layers and gradient through variants of backpropagation. However, it is well understood that gradients are noisy and alternatives like guided backpropagation have been proposed to obtain better visualization at inference. In this work, we present a novel training approach to improve the quality of gradients for interpretability. In particular, we introduce a regularization loss such that the gradient with respect to the input image obtained by standard backpropagation is similar to the gradient obtained by guided backpropagation. We find that the resulting gradient is qualitatively less noisy and improves quantitatively the interpretability properties of different networks, using several interpretability methods.

Explanations obtained from transformer-based architectures in the form of raw attention, can be seen as a class-agnostic saliency map. Additionally, attention-based pooling serves as a form of masking the in feature space. Motivated by this observation, we design an attention-based pooling mechanism intended to replace Global Average Pooling (GAP) at inference. This mechanism, called Cross-Attention Stream (CA-Stream), comprises a stream of cross attention blocks interacting with features at different network depths. CA-Stream enhances interpretability in models, while preserving recognition performance.

How important is it for training and evaluation sets to not have class overlap in image retrieval? We revisit Google Landmarks v2 clean, the most popular training set, by identifying and removing class overlap with Revisited Oxford and Paris [34], the most popular evaluation set. By comparing the original and the new RGLDv2-clean on a benchmark of reproduced state-of-the-art methods, our findings are striking. Not only is there a dramatic drop in performance, but it is inconsistent across methods, changing the ranking.What does it take to focus on objects or interest and ignore background clutter when indexing? Do we need to train an object detector and the representation separately? Do we need location supervision? We introduce Single-stage Detect-to-Retrieve (CiDeR), an end-to-end, single-stage pipeline to detect objects of interest and extract a global image representation. We outperform previous state-of-the-art on both existing training sets and the new RGLDv2-clean. Our dataset is available at this https URL.

Multimodal sentiment analysis (MSA) leverages heterogeneous data sources to interpret the complex nature of human sentiments. Despite significant progress in multimodal architecture design, the field lacks comprehensive regularization methods. This paper introduces PowMix, a versatile embedding space regularizer that builds upon the strengths of unimodal mixing-based regularization approaches and introduces novel algorithmic components that are specifically tailored to multimodal tasks. PowMix is integrated before the fusion stage of multimodal architectures and facilitates intra-modal mixing, such as mixing text with text, to act as a regularizer. PowMix consists of five components: 1) a varying number of generated mixed examples, 2) mixing factor reweighting, 3) anisotropic mixing, 4) dynamic mixing, and 5) cross-modal label mixing. Extensive experimentation across benchmark MSA datasets and a broad spectrum of diverse architectural designs demonstrate the efficacy of PowMix, as evidenced by consistent performance improvements over baselines and existing mixing methods. An in-depth ablation study highlights the critical contribution of each PowMix component and how they synergistically enhance performance. Furthermore, algorithmic analysis demonstrates how PowMix behaves in different scenarios, particularly comparing early versus late fusion architectures. Notably, PowMix enhances overall performance without sacrificing model robustness or magnifying text dominance. It also retains its strong performance in situations of limited data. Our findings position PowMix as a promising versatile regularization strategy for MSA. Code will be made available.

Mixup refers to interpolation-based data augmentation, originally motivated as a way to go beyond empirical risk minimization (ERM). Its extensions mostly focus on the definition of interpolation and the space (input or feature) where it takes place, while the augmentation process itself is less studied. In most methods, the number of generated examples is limited to the mini-batch size and the number of examples being interpolated is limited to two (pairs), in the input space.

We make progress in this direction by introducing MultiMix, which generates an arbitrarily large number of interpolated examples beyond the mini-batch size and interpolates the entire mini-batch in the embedding space. Effectively, we sample on the entire convex hull of the mini-batch rather than along linear segments between pairs of examples.

On sequence data, we further extend to Dense MultiMix. We densely interpolate features and target labels at each spatial location and also apply the loss densely. To mitigate the lack of dense labels, we inherit labels from examples and weight interpolation factors by attention as a measure of confidence.

Overall, we increase the number of loss terms per mini-batch by orders of magnitude at little additional cost. This is only possible because of interpolating in the embedding space. We empirically show that our solutions yield significant improvement over state-of-the-art mixup methods on four different benchmarks, despite interpolation being only linear. By analyzing the embedding space, we show that the classes are more tightly clustered and uniformly spread over the embedding space, thereby explaining the improved behavior.

Few-shot learning addresses the issue of classifying images using limited labeled data. Exploiting unlabeled data through the use of transductive inference methods such as label propagation has been shown to improve the performance of few-shot learning significantly. Label propagation infers pseudo-labels for unlabeled data by utilizing a constructed graph that exploits the underlying manifold structure of the data. However, a limitation of the existing label propagation approaches is that the positions of all data points are fixed and might be sub-optimal so that the algorithm is not as effective as possible. In this work, we propose a novel algorithm that adapts the feature embeddings of the labeled data by minimizing a differentiable loss function optimizing their positions in the manifold in the process. Our novel algorithm, Adaptive Anchor Label Propagation}, outperforms the standard label propagation algorithm by as much as 7% and 2% in the 1-shot and 5-shot settings respectively. We provide experimental results highlighting the merits of our algorithm on four widely used few-shot benchmark datasets, namely miniImageNet, tieredImageNet, CUB and CIFAR-FS and two commonly used backbones, ResNet12 and WideResNet-28-10. The source code can be found at this https URL.

Self-supervised learning has unlocked the potential of scaling up pretraining to billions of images, since annotation is unnecessary. But are we making the best use of data? How more economical can we be? In this work, we attempt to answer this question by making two contributions. First, we investigate first-person videos and introduce a "Walking Tours" dataset. These videos are high-resolution, hours-long, captured in a single uninterrupted take, depicting a large number of objects and actions with natural scene transitions. They are unlabeled and uncurated, thus realistic for self-supervision and comparable with human learning.

Second, we introduce a novel self-supervised image pretraining method tailored for learning from continuous videos. Existing methods typically adapt image-based pretraining approaches to incorporate more frames. Instead, we advocate a "tracking to learn to recognize" approach. Our method called DoRA, leads to attention maps that Discover and tRAck objects over time in an end-to-end manner, using transformer cross-attention. We derive multiple views from the tracks and use them in a classical self-supervised distillation loss. Using our novel approach, a single Walking Tours video remarkably becomes a strong competitor to ImageNet for several image and video downstream tasks.

Recent vision-language models are driven by large-scale pretrained models. However, adapting pretrained models on limited data presents challenges such as overfitting, catastrophic forgetting, and the cross-modal gap between vision and language. We introduce a parameter-efficient method to address these challenges, combining multimodal prompt learning and a transformer-based mapping network, while keeping the pretrained models frozen. Our experiments on several video question answering benchmarks demonstrate the superiority of our approach in terms of performance and parameter efficiency on both zero-shot and few-shot settings. Our code is available at this https URL.

Convolutional networks and vision transformers have different forms of pairwise interactions, pooling across layers and pooling at the end of the network. Does the latter really need to be different? As a by-product of pooling, vision transformers provide spatial attention for free, but this is most often of low quality unless self-supervised, which is not well studied. Is supervision really the problem?

In this work, we develop a generic pooling framework and then we formulate a number of existing methods as instantiations. By discussing the properties of each group of methods, we derive SimPool, a simple attention-based pooling mechanism as a replacement of the default one for both convolutional and transformer encoders. We find that, whether supervised or self-supervised, this improves performance on pre-training and downstream tasks and provides attention maps delineating object boundaries in all cases. One could thus call SimPool universal. To our knowledge, we are the first to obtain attention maps in supervised transformers of at least as good quality as self-supervised, without explicit losses or modifying the architecture. Code at: this https URL.

Transductive few-shot learning algorithms have showed substantially superior performance over their inductive counterparts by leveraging the unlabeled queries. However, the vast majority of such methods are evaluated on perfectly class-balanced benchmarks. It has been shown that they undergo remarkable drop in performance under a more realistic, imbalanced setting. To this end, we propose a novel algorithm to address imbalanced transductive few-shot learning, named Adaptive Manifold. Our method exploits the underlying manifold of the labeled support examples and unlabeled queries by using manifold similarity to predict the class probability distribution per query. It is parameterized by one centroid per class as well as a set of graph-specific parameters that determine the manifold. All parameters are optimized through a loss function that can be tuned towards class-balanced or imbalanced distributions. The manifold similarity shows substantial improvement over Euclidean distance, especially in the 1-shot setting. Our algorithm outperforms or is on par with other state of the art methods in three benchmark datasets, namely miniImageNet, tieredImageNet and CUB, and three different backbones, namely ResNet-18, WideResNet-28-10 and DenseNet-121. In certain cases, our algorithm outperforms the previous state of the art by as much as 4.2%.

Impressive progress in generative models and implicit representations gave rise to methods that can generate 3D shapes of high quality. However, being able to locally control and edit shapes is another essential property that can unlock several content creation applications. Local control can be achieved with part-aware models, but existing methods require 3D supervision and cannot produce textures. In this work, we devise PartNeRF, a novel part-aware generative model for editable 3D shape synthesis that does not require any explicit 3D supervision. Our model generates objects as a set of locally defined NeRFs, augmented with an affine transformation. This enables several editing operations such as applying transformations on parts, mixing parts from different objects etc. To ensure distinct, manipulable parts we enforce a hard assignment of rays to parts that makes sure that the color of each ray is only determined by a single NeRF. As a result, altering one part does not affect the appearance of the others. Evaluations on various ShapeNet categories demonstrate the ability of our model to generate editable 3D objects of improved fidelity, compared to previous part-based generative approaches that require 3D supervision or models relying on NeRFs.

Methods based on class activation maps (CAM) provide a simple mechanism to interpret predictions of convolutional neural networks by using linear combinations of feature maps as saliency maps. By contrast, masking-based methods optimize a saliency map directly in the image space or learn it by training another network on additional data.

In this work we introduce Opti-CAM, combining ideas from CAM-based and masking-based approaches. Our saliency map is a linear combination of feature maps, where weights are optimized per image such that the logit of the masked image for a given class is maximized. We also fix a fundamental flaw in two of the most common evaluation metrics of attribution methods. On several datasets, Opti-CAM largely outperforms other CAM-based approaches according to the most relevant classification metrics. We provide empirical evidence supporting that localization and classifier interpretability are not necessarily aligned.

Vision transformers have achieved remarkable progress in vision tasks such as image classification and detection. However, in instance-level image retrieval, transformers have not yet shown good performance compared to convolutional networks. We propose a number of improvements that make transformers outperform the state of the art for the first time. (1) We show that a hybrid architecture is more effective than plain transformers, by a large margin. (2) We introduce two branches collecting global (classification token) and local (patch tokens) information, from which we form a global image representation. (3) In each branch, we collect multi-layer features from the transformer encoder, corresponding to skip connections across distant layers. (4) We enhance locality of interactions at the deeper layers of the encoder, which is the relative weakness of vision transformers. We train our model on all commonly used training sets and, for the first time, we make fair comparisons separately per training set. In all cases, we outperform previous models based on global representation. Public code is available at this https URL.

Mixup refers to interpolation-based data augmentation, originally motivated as a way to go beyond empirical risk minimization (ERM). Yet, its extensions focus on the definition of interpolation and the space where it takes place, while the augmentation itself is less studied: For a mini-batch of size $m$, most methods interpolate between $m$ pairs with a single scalar interpolation factor $\lambda$.

In this work, we make progress in this direction by introducing MultiMix, which interpolates an arbitrary number $n$ of tuples, each of length $m$, with one vector $\lambda$ per tuple. On sequence data, we further extend to dense interpolation and loss computation over all spatial positions. Overall, we increase the number of tuples per mini-batch by orders of magnitude at little additional cost. This is possible by interpolating at the very last layer before the classifier. Finally, to address inconsistencies due to linear target interpolation, we introduce a self-distillation approach to generate and interpolate synthetic targets.

We empirically show that our contributions result in significant improvement over state-of-the-art mixup methods on four benchmarks. By analyzing the embedding space, we observe that the classes are more tightly clustered and uniformly spread over the embedding space, thereby explaining the improved behavior.

Transformers and masked language modeling are quickly being adopted and explored in computer vision as vision transformers and masked image modeling (MIM). In this work, we argue that image token masking differs from token masking in text, due to the amount and correlation of tokens in an image. In particular, to generate a challenging pretext task for MIM, we advocate a shift from random masking to informed masking. We develop and exhibit this idea in the context of distillation-based MIM, where a teacher transformer encoder generates an attention map, which we use to guide masking for the student. We thus introduce a novel masking strategy, called attention-guided masking (AttMask), and we demonstrate its effectiveness over random masking for dense distillation-based MIM as well as plain distillation-based self-supervised learning on classification tokens. We confirm that AttMask accelerates the learning process and improves the performance on a variety of downstream tasks. We provide the implementation code at this https URL.

We address representation learning for large-scale instance-level image retrieval. Apart from backbone, training pipelines and loss functions, popular approaches have focused on different spatial pooling and attention mechanisms, which are at the core of learning a powerful global image representation. There are different forms of attention according to the interaction of elements of the feature tensor (local and global) and the dimensions where it is applied (spatial and channel). Unfortunately, each study addresses only one or two forms of attention and applies it to different problems like classification, detection or retrieval.

We present global-local attention module (GLAM), which is attached at the end of a backbone network and incorporates all four forms of attention: local and global, spatial and channel. We obtain a new feature tensor and, by spatial pooling, we learn a powerful embedding for image retrieval. Focusing on global descriptors, we provide empirical evidence of the interaction of all forms of attention and improve the state of the art on standard benchmarks.

Few-shot classification addresses the challenge of classifying examples given not just limited supervision but limited data as well. An attractive solution is synthetic data generation. However, most such methods are overly sophisticated, focusing on high-quality, realistic data in the input space. It is unclear whether adapting them to the few-shot regime and using them for the downstream task of classification is the right approach. Previous works on synthetic data generation for few-shot classification focus on exploiting complex models, e.g. a Wasserstein GAN with multiple regularizers or a network that transfers latent diversities from known to novel classes.

We follow a different approach and investigate how a simple and straightforward synthetic data generation method can be used effectively. We make two contributions, namely we show that: (1) using a simple loss function is more than enough for training a feature generator in the few-shot setting; and (2) learning to generate tensor features instead of vector features is superior. Extensive experiments on miniImagenet, CUB and CIFAR-FS datasets show that our method sets a new state of the art, outperforming more sophisticated few-shot data augmentation methods.

Metric learning involves learning a discriminative representation such that embeddings of similar classes are encouraged to be close, while embeddings of dissimilar classes are pushed far apart. State-of-the-art methods focus mostly on sophisticated loss functions or mining strategies. On the one hand, metric learning losses consider two or more examples at a time. On the other hand, modern data augmentation methods for classification consider two or more examples at a time. The combination of the two ideas is under-studied.

In this work, we aim to bridge this gap and improve representations using mixup, which is a powerful data augmentation approach interpolating two or more examples and corresponding target labels at a time. This task is challenging because, unlike classification, the loss functions used in metric learning are not additive over examples, so the idea of interpolating target labels is not straightforward. To the best of our knowledge, we are the first to investigate mixing examples and target labels for deep metric learning. We develop a generalized formulation that encompasses existing metric learning loss functions and modify it to accommodate for mixup, introducing Metric Mix, or Metrix. We show that mixing inputs, intermediate representations or embeddings along with target labels significantly improves representations and outperforms state-of-the-art metric learning methods on four benchmark datasets.

Few-shot classification addresses the challenge of classifying examples given only limited labeled data. A powerful approach is to go beyond data augmentation, towards data synthesis. However, most of data augmentation/synthesis methods for few-shot classification are overly complex and sophisticated, e.g. training a wGAN with multiple regularizers or training a network to transfer latent diversities from known to novel classes. We make two contributions, namely we show that: (1) using a simple loss function is more than enough for training a feature generator in the few-shot setting; and (2) learning to generate tensor features instead of vector features is superior. Extensive experiments on miniImagenet, CUB and CIFAR-FS datasets show that our method sets a new state of the art, outperforming more sophisticated few-shot data augmentation methods.

Mixup is a powerful data augmentation method that interpolates between two or more examples in the input or feature space and between the corresponding target labels. Many recent mixup methods focus on cutting and pasting two or more objects into one image, which is more about efficient processing than interpolation. However, how to best interpolate images is not well defined. In this sense, mixup has been connected to autoencoders, because often autoencoders "interpolate well", for instance generating an image that continuously deforms into another.

In this work, we revisit mixup from the interpolation perspective and introduce AlignMix, where we geometrically align two images in the feature space. The correspondences allow us to interpolate between two sets of features, while keeping the locations of one set. Interestingly, this gives rise to a situation where mixup retains mostly the geometry or pose of one image and the texture of the other, connecting it to style transfer. More than that, we show that an autoencoder can still improve representation learning under mixup, without the classifier ever seeing decoded images. AlignMix outperforms state-of-the-art mixup methods on five different benchmarks.

High-level understanding of stories in video such as movies and TV shows from raw data is extremely challenging. Modern video question answering (VideoQA) systems often use additional human-made sources like plot synopses, scripts, video descriptions or knowledge bases. In this work, we present a new approach to understand the whole story without such external sources. The secret lies in the dialog: unlike any prior work, we treat dialog as a noisy source to be converted into text description via dialog summarization, much like recent methods treat video. The input of each modality is encoded by transformers independently, and a simple fusion method combines all modalities, using soft temporal attention for localization over long inputs. Our model outperforms the state of the art on the KnowIT VQA dataset by a large margin, without using question-specific human annotation or human-made plot summaries. It even outperforms human evaluators who have never watched any whole episode before.

The challenge in few-shot learning is that available data is not enough to capture the underlying distribution. To mitigate this, two emerging directions are (a) using local image representations, essentially multiplying the amount of data by a constant factor, and (b) using more unlabeled data, for instance by transductive inference, jointly on a number of queries. In this work, we bring these two ideas together, introducing local propagation. We treat local image features as independent examples, we build a graph on them and we use it to propagate both the features themselves and the labels, known and unknown. Interestingly, since there is a number of features per image, even a single query gives rise to transductive inference. As a result, we provide a universally safe choice for few-shot inference under both non-transductive and transductive settings, improving accuracy over corresponding methods. This is in contrast to existing solutions, where one needs to choose the method depending on the quantity of available data.

Few-shot learning amounts to learning representations and acquiring knowledge such that novel tasks may be solved with both supervision and data being limited. Improved performance is possible by transductive inference, where the entire test set is available concurrently, and semi-supervised learning, where more unlabeled data is available. These problems are closely related because there is little or no adaptation of the representation in novel tasks.

Focusing on these two settings, we introduce a new algorithm that leverages the manifold structure of the labeled and unlabeled data distribution to predict pseudo-labels, while balancing over classes and using the loss value distribution of a limited-capacity classifier to select the cleanest labels, iterately improving the quality of pseudo-labels. Our solution sets new state of the art on four benchmark datasets, namely miniImageNet, tieredImageNet, CUB and CIFAR-FS, while being robust over feature space pre-processing and the quantity of available data.

Knowledge transfer from large teacher models to smaller student models has recently been studied for metric learning, focusing on fine-grained classification. In this work, focusing on instance-level image retrieval, we study an asymmetric testing task, where the database is represented by the teacher and queries by the student. Inspired by this task, we introduce asymmetric metric learning, a novel paradigm of using asymmetric representations at training. This acts as a simple combination of knowledge transfer with the original metric learning task.

We systematically evaluate different teacher and student models, metric learning and knowledge transfer loss functions on the new asymmetric testing as well as the standard symmetric testing task, where database and queries are represented by the same model. We find that plain regression is surprisingly effective compared to more complex knowledge transfer mechanisms, working best in asymmetric testing. Interestingly, our asymmetric metric learning approach works best in symmetric testing, allowing the student to even outperform the teacher.

Few-shot learning is often motivated by the ability of humans to learn new tasks from few examples. However, standard few-shot classification benchmarks assume that the representation is learned on a limited amount of base class data, ignoring the amount of prior knowledge that a human may have accumulated before learning new tasks. At the same time, even if a powerful representation is available, it may happen in some domain that base class data are limited or non-existent. This motivates us to study a problem where the representation is obtained from a classifier pre-trained on a large-scale dataset of a different domain, assuming no access to its training process, while the base class data are limited to few examples per class and their role is to adapt the representation to the domain at hand rather than learn from scratch. We adapt the representation in two stages, namely on the few base class data if available and on the even fewer data of new tasks. In doing so, we obtain from the pre-trained classifier a spatial attention map that allows focusing on objects and suppressing background clutter. This is important in the new problem, because when base class data are few, the network cannot learn where to focus implicitly. We also show that a pre-trained network may be easily adapted to novel classes, without meta-learning.

Adversarial examples of deep neural networks are receiving ever increasing attention because they help in understanding and reducing the sensitivity to their input. This is natural given the increasing applications of deep neural networks in our everyday lives. When white-box attacks are almost always successful, it is typically only the distortion of the perturbations that matters in their evaluation.

In this work, we argue that speed is important as well, especially when considering that fast attacks are required by adversarial training. Given more time, iterative methods can always find better solutions. We investigate this speed-distortion trade-off in some depth and introduce a new attack called boundary projection (BP) that improves upon existing methods by a large margin. Our key idea is that the classification boundary is a manifold in the image space: we therefore quickly reach the boundary and then optimize distortion on this manifold.

Weakly-supervised object detection attempts to limit the amount of supervision by dispensing the need for bounding boxes, but still assumes image-level labels on the entire training set are available. In this work, we study the problem of training an object detector from one or few clean images with image-level labels and a larger set of completely unlabeled images. This is an extreme case of semi-supervised learning where the labeled data are not enough to bootstrap the learning of a classifier or detector. Our solution is to use a standard weakly-supervised pipeline to train a student model from image-level pseudo-labels generated on the unlabeled set by a teacher model, bootstrapped by region-level similarities to clean labeled images. By using the recent pipeline of PCL and more unlabeled images, we achieve performance competitive or superior to many state of the art weakly-supervised detection solutions.

Active learning typically focuses on training a model on few labeled examples alone, while unlabeled ones are only used for acquisition. In this work we depart from this setting by using both labeled and unlabeled data during model training across active learning cycles. We do so by using unsupervised feature learning at the beginning of the active learning pipeline and semi-supervised learning at every active learning cycle, on all available data. The former has not been investigated before in active learning, while the study of latter in the context of deep learning is scarce and recent findings are not conclusive with respect to its benefit. Our idea is orthogonal to acquisition strategies by using more data, much like ensemble methods use more models. By systematically evaluating on a number of popular acquisition strategies and datasets, we find that the use of unlabeled data during model training brings a spectacular accuracy improvement in image classification, compared to the differences between acquisition strategies. We thus explore smaller label budgets, even one label per class.

In this work we consider the problem of learning a classifier from noisy labels when a few clean labeled examples are given. The structure of clean and noisy data is modeled by a graph per class and Graph Convolutional Networks (GCN) are used to predict class relevance of noisy examples. For each class, the GCN is treated as a binary classifier learning to discriminate clean from noisy examples using a weighted binary cross-entropy loss function, and then the GCN-inferred "clean" probability is exploited as a relevance measure. Each noisy example is weighted by its relevance when learning a classifier for the end task. We evaluate our method on an extended version of a few-shot learning problem, where the few clean examples of novel classes are supplemented with additional noisy data. Experimental results show that our GCN-based cleaning process significantly improves the classification accuracy over not cleaning the noisy data and standard few-shot classification where only few clean examples are used. The proposed GCN-based method outperforms the transductive approach (Douze et al., 2018) that is using the same additional data without labels.

We propose a novel method of deep spatial matching (DSM) for image retrieval. Initial ranking is based on image descriptors extracted from convolutional neural network activations by global pooling, as in recent state-of-the-art work. However, the same sparse 3D activation tensor is also approximated by a collection of local features. These local features are then robustly matched to approximate the optimal alignment of the tensors. This happens without any network modification, additional layers or training. No local feature detection happens on the original image. No local feature descriptors and no visual vocabulary are needed throughout the whole process.

We experimentally show that the proposed method achieves the state-of-the-art performance on standard benchmarks across different network architectures and different global pooling methods. The highest gain in performance is achieved when diffusion on the nearest-neighbor graph of global descriptors is initiated from spatially verified images.

Semi-supervised learning is becoming increasingly important because it can combine data carefully labeled by humans with abundant unlabeled data to train deep neural networks. Classic methods on semi-supervised learning that have focused on transductive learning have not been fully exploited in the inductive framework followed by modern deep learning. The same holds for the manifold assumption---that similar examples should get the same prediction. In this work, we employ a transductive label propagation method that is based on the manifold assumption to make predictions on the entire dataset and use these predictions to generate pseudo-labels for the unlabeled data and train a deep neural network. At the core of the transductive method lies a nearest neighbor graph of the dataset that we create based on the embeddings of the same network.Therefore our learning process iterates between these two steps. We improve performance on several datasets especially in the few labels regime and show that our work is complementary to current state of the art.

This paper investigates the visual quality of the adversarial examples. Recent papers propose to smooth the perturbations to get rid of high frequency artefacts. In this work, smoothing has a different meaning as it perceptually shapes the perturbation according to the visual content of the image to be attacked. The perturbation becomes locally smooth on the flat areas of the input image, but it may be noisy on its textured areas and sharp across its edges.

This operation relies on Laplacian smoothing, well-known in graph signal processing, which we integrate in the attack pipeline. We benchmark several attacks with and without smoothing under a white-box scenario and evaluate their transferability. Despite the additional constraint of smoothness, our attack has the same probability of success at lower distortion.

Training deep neural networks from few examples is a highly challenging and key problem for many computer vision tasks. In this context, we are targeting knowledge transfer from a set with abundant data to other sets with few available examples. We propose two simple and effective solutions: (i) dense classification over feature maps, which for the first time studies local activations in the domain of few-shot learning, and (ii) implanting, that is, attaching new neurons to a previously trained network to learn new, task-specific features. On miniImageNet, we improve the prior state-of-the-art on few-shot classification, i.e., we achieve 62.5%, 79.8% and 83.8% on 5-way 1-shot, 5-shot and 10-shot settings respectively.

State of the art image retrieval performance is achieved with CNN features and manifold ranking using a k-NN similarity graph that is pre-computed off-line. The two most successful existing approaches are temporal filtering, where manifold ranking amounts to solving a sparse linear system online, and spectral filtering, where eigen-decomposition of the adjacency matrix is performed off-line and then manifold ranking amounts to dot-product search online. The former suffers from expensive queries and the latter from significant space overhead. Here we introduce a novel, theoretically well-founded hybrid filtering approach allowing full control of the space-time trade-off between these two extremes. Experimentally, we verify that our hybrid method delivers results on par with the state of the art, with lower memory demands compared to spectral filtering approaches and faster compared to temporal filtering.

In this paper we address issues with image retrieval benchmarking on standard and popular Oxford 5k and Paris 6k datasets. In particular, annotation errors, the size of the dataset, and the level of challenge are addressed: new annotation for both datasets is created with an extra attention to the reliability of the ground truth. Three new protocols of varying difficulty are introduced. The protocols allow fair comparison between different methods, including those using a dataset pre-processing stage. For each dataset, 15 new challenging queries are introduced. Finally, a new set of 1M hard, semi-automatically cleaned distractors is selected.

An extensive comparison of the state-of-the-art methods is performed on the new benchmark. Different types of methods are evaluated, ranging from local-feature-based to modern CNN based methods. The best results are achieved by taking the best of the two worlds. Most importantly, image retrieval appears far from being solved.