Yannis Avrithis
Research Scientist
Inria Rennes-Bretagne Atlantique, France

Working on computer vision and machine learning

# News

May 15, 2019
Research grant GRAPES accepted
Apr 22, 2019
I am a Senior Member of IEEE
Apr 10, 2019
Technical report released at arXiv: Semi-Supervised Learning
Mar 29, 2019
Technical report released at arXiv: Smooth Adversarial Examples
Mar 13, 2019
Technical report released at arXiv: Few-Shot Learning
Mar 07, 2019
Feb 26, 2019
New website online at avrithis.net, replacing IVA
Feb 25, 2019
Three papers accepted at CVPR 2019: Deep Spatial Matching, Few-Shot Learning and Semi-Supervised Learning
Feb 19, 2019
I begin teaching the 1st iteration of Computer Vision at NKUA
Feb 08, 2019
Paper accepted at MVA: Particular Object Discovery
Nov 20, 2018
Paper accepted at CVIU: Planar Shape Decomposition
Nov 19, 2018
I begin teaching the 2nd iteration of Deep Learning for Vision at UR1
Oct 10, 2018
Research grant BnF begins
Sep 21, 2018
Paper accepted at ACCV 2018: Hybrid Diffusion
Sep 01, 2018
Research grant MobilAI begins
Jul 23, 2018
Technical report released at arXiv: Hybrid Diffusion
May 10, 2018
Mar 29, 2018
Technical report released at arXiv: Revisiting Oxford and Paris
Mar 29, 2018
Technical report released at arXiv: Mining on Manifolds
Mar 22, 2018
I give a talk on Deep Image Retrieval at Safran Tech
Feb 28, 2018
Three papers accepted at CVPR 2018: Fast Spectral Ranking, Revisiting Oxford and Paris and Mining on Manifolds
Feb 19, 2018
Yann Lifchitz begins his PhD thesis
Feb 19, 2018
Research grant Few-Shot begins
Jan 19, 2018
Paper accepted at WACV 2018: Unsupervised Object Discovery

Since 2016 I am a research scientist in LinkMedia team of Inria Rennes-Bretagne Atlantique, carrying out research on computer vision and machine learning, and teaching Deep Learning for Vision. I enjoy working on learning visual representations from data, with as little supervision as possible. My recent work is focusing on exploring the manifold structure of data and, apart from image retrieval, using it for unsupervised, semi-supervised and few-shot learning. I am also working on adversarial examples and on investigating the sparsity of convolutional activations, applied to spatial matching and unsupervised object discovery.

In 2015 I have been at the Laboratory of Algebraic and Geometric Algorithms (ΕρΓΑ) of National and Kapodistrian University of Athens (NKUA), where I have worked on large-scale clustering and nearest neighbor search with Ioannis Emiris. An achievement of this period has been Web-Scale Image Clustering.

Between 2001 and 2015 I have been at the Image, Video and Multimedia Systems Laboratory (IVML) of the National Technical University of Athens (NTUA). In the latter part of this period, since 2008, I have been leading the Image and Video Analysis (IVA) team. We have worked on a diverse set of problems including local feature and salient region detection, spatial and spatiotemporal visual representation and matching, object recognition and tracking, action recognition in video, scene classification, image/video indexing, retrieval and summarization.

In 2015 I have been at the Laboratory of Algebraic and Geometric Algorithms (ΕρΓΑ) of National and Kapodistrian University of Athens (NKUA), where I have worked on large-scale clustering and nearest neighbor search with Ioannis Emiris. An achievement of this period has been Web-Scale Image Clustering.

Between 2001 and 2015 I have been at the Image, Video and Multimedia Systems Laboratory (IVML) of the National Technical University of Athens (NTUA). In the latter part of this period, since 2008, I have been leading the Image and Video Analysis (IVA) team. We have worked on a diverse set of problems including local feature and salient region detection, spatial and spatiotemporal visual representation and matching, object recognition and tracking, action recognition in video, scene classification, image/video indexing, retrieval and summarization.

A large body of this work has focused on local feature/descriptor representations. In this context, we have delivered repeatable feature detection, extremely fast spatial matching, geometry indexing, multi-view and single-view feature selection, approximate nearest neighbor search for large scale clustering and vocabulary construction, and mining 3d scenes from millions of images. Examples of this work are Hough Pyramid Matching (HPM), the Aggregated Selective Match Kernel (ASMK), and Locally Optimized Product Quantization (LOPQ). In 2017, using a CNN image representation, Yahoo! Research has chosen LOPQ to index and provide a similar image search functionality on its entire Flickr collection. A flagship product of this period has been our unique application Visual Image Retrieval and Localization (VIRaL).

In the preceding period before 2008, I have worked on problems related to object detection and image understanding. This includes regions-of-interest for generic object detection, semantic segmentation, integrated segmentation and region labeling, spatio-temporal object segmentation and tracking, image classification, as well as the use of visual context and prior knowledge.

While at NTUA, I have had the opportunity to collaborate with a large number of researchers across Europe by participating in Networks of Excellence Muscle and K-Space and Integrated Projects like aceMedia and WeKnowIt. I have developed a number of activities including being the Program Chair or General Chair of workshops and conferences in the field of multimedia like WIAMIS, CBMI, MMM and CIVR. For several years I have been teaching Signals and Systems and Image and Video Analysis.

Between 1996 and 2001 I have been working on my Ph.D. at NTUA with Stefanos Kollias, studying visual representations for video sequence analysis. I have investigated accurate spatio-temporal segmentation by fusing color, motion and depth, as well as a region-based representation for retrieval and summarization in the form of key-frames. I have designed an affine-invariant representation of object contours for shape-based object classification and retrieval. Finally, I have studied temporal segmentation and parsing of broadcast news based on human face detection.

In 1993-1994 I completed a M.Sc. with Distinction in Communications and Signal Processing at Imperial College, University of London. As part of my Masters thesis, I have worked on communications and investigated the capacity of a cellular CDMA system with Athanassios Manikas.

Between 1998 and 1993 I have studied Electrical and Computer Engineering at NTUA. As part of my Diploma thesis, I have worked on analog electronics and developed a hardware implementation of a fuzzy logic processor with Yannis Tsividis. Before that, my first exposure to computer vision and machine learning has been the study of an invariant representation for optical character recognition with Anastasios Delopoulos, leading to a conference publication before my Diploma.

# Recent publications

Label Propagation for Deep Semi-Supervised Learning
CVPR 2019
Dense Classification and Implanting for Few-Shot Learning
CVPR 2019
Local Features and Visual Words Emerge in Activations
CVPR 2019
Graph-Based Particular Object Discovery
Revisiting the Medial Axis for Planar Shape Decomposition
Hybrid Diffusion: Spectral-Temporal Graph Filtering for Manifold Ranking
ACCV 2018
Mining on Manifolds: Metric Learning without Labels
Revisiting Oxford And Paris: Large-Scale Image Retrieval Benchmarking
Fast Spectral Ranking for Similarity Search
Unsupervised Object Discovery for Instance Recognition
Automatic Discovery of Discriminative Parts as a Quadratic Assignment Problem
Unsupervised Part Learning for Visual Recognition
Panorama to Panorama Matching for Location Recognition
$\alpha$-Shapes for Local Feature Detection
Web-Scale Image Clustering Revisited
Planar Shape Decomposition Made Simple
Dithering-Based Sampling and Weighted $\alpha$-Shapes for Local Feature Detection
Improving Local Features by Dithering-Based Image Sampling
Towards Large-Scale Geometry Indexing by Feature Selection