Laboratoire d'Informatique Gaspard-Monge

Deep residual networks were shown to be able to scale up to thousands of layers and still have improving performance. However, each fraction of a percent of improved accuracy costs nearly doubling the number of layers, and so training very deep residual networks has a problem of diminishing feature reuse, which makes these networks very slow to train. To tackle these problems, in this paper we conduct a detailed experimental study on the architecture of ResNet blocks, based on which we propose a novel architecture where we decrease depth and increase width of residual networks. We call the resulting network structures wide residual networks (WRNs) and show that these are far superior over their commonly used thin and very deep counterparts. For example, we demonstrate that even a simple 16-layer-deep wide residual network outperforms in accuracy and efficiency all previous deep residual networks, including thousand-layer-deep networks, achieving new state-of-the-art results on CIFAR, SVHN, COCO, and significant improvements on ImageNet. Our code and models are available at this https URL

Scene labeling consists of labeling each pixel in an image with the category of the object it belongs to. We propose a method that uses a multiscale convolutional network trained from raw pixels to extract dense feature vectors that encode regions of multiple sizes centered on each pixel. The method alleviates the need for engineered features, and produces a powerful representation that captures texture, shape, and contextual information. We report results using multiple postprocessing methods to produce the final labeling. Among those, we propose a technique to automatically retrieve, from a pool of segmentation components, an optimal set of components that best explain the scene; these components are arbitrary, for example, they can be taken from a segmentation tree or from any family of oversegmentations. The system yields record accuracies on the SIFT Flow dataset (33 classes) and the Barcelona dataset (170 classes) and near-record accuracy on Stanford background dataset (eight classes), while being an order of magnitude faster than competing approaches, producing a $(320\times 240)$ image labeling in less than a second, including feature extraction.

Deep residual networks were shown to be able to scale up to thousands of layers and still have improving performance. However, each fraction of a percent of improved accuracy costs nearly doubling the number of layers, and so training very deep residual networks has a problem of diminishing feature reuse, which makes these networks very slow to train. To tackle these problems, in this paper we conduct a detailed experimental study on the architecture of ResNet blocks, based on which we propose a novel architecture where we decrease depth and increase width of residual networks. We call the resulting network structures wide residual networks (WRNs) and show that these are far superior over their commonly used thin and very deep counterparts. For example, we demonstrate that even a simple 16-layer-deep wide residual network outperforms in accuracy and efficiency all previous deep residual networks, including thousand-layer-deep networks, achieving new state-of-the-art results on CIFAR, SVHN, COCO, and significant improvements on ImageNet. Our code and models are available at https://github.com/szagoruyko/wide-residual-networks

Monocular depth estimation, which plays a crucial role in understanding 3D scene geometry, is an ill-posed problem. Recent methods have gained significant improvement by exploring image-level information and hierarchical features from deep convolutional neural networks (DCNNs). These methods model depth estimation as a regression problem and train the regression networks by minimizing mean squared error, which suffers from slow convergence and unsatisfactory local solutions. Besides, existing depth estimation networks employ repeated spatial pooling operations, resulting in undesirable low-resolution feature maps. To obtain high-resolution depth maps, skip-connections or multilayer deconvolution networks are required, which complicates network training and consumes much more computations. To eliminate or at least largely reduce these problems, we introduce a spacing-increasing discretization (SID) strategy to discretize depth and recast depth network learning as an ordinal regression problem. By training the network using an ordinary regression loss, our method achieves much higher accuracy and faster convergence in synch. Furthermore, we adopt a multi-scale network structure which avoids unnecessary spatial pooling and captures multi-scale information in parallel. The proposed deep ordinal regression network (DORN) achieves state-of-the-art results on three challenging benchmarks, i.e., KITTI [16], Make3D [49], and NYU Depth v2 [41], and outperforms existing methods by a large margin.

Combinatorics on words has arisen independently within several branches of mathematics, for instance number theory, group theory and probability, and appears frequently in problems related to theoretical computer science. The first unified treatment of the area was given in Lothaire's book Combinatorics on Words. Originally published in 2002, this book presents several more topics and provides deeper insights into subjects discussed in the previous volume. An introductory chapter provides the reader with all the necessary background material. There are numerous examples, full proofs whenever possible and a notes section discussing further developments in the area. This book is both a comprehensive introduction to the subject and a valuable reference source for researchers.

This paper presents a method to automatically and efficiently detect face tampering in videos, and particularly focuses on two recent techniques used to generate hyper-realistic forged videos: Deepfake and Face2Face. Traditional image forensics techniques are usually not well suited to videos due to the compression that strongly degrades the data. Thus, this paper follows a deep learning approach and presents two networks, both with a low number of layers to focus on the mesoscopic properties of images. We evaluate those fast networks on both an existing dataset and a dataset we have constituted from online videos. The tests demonstrate a very successful detection rate with more than 98% for Deepfake and 95% for Face2Face.

Attention plays a critical role in human visual experience. Furthermore, it has recently been demonstrated that attention can also play an important role in the context of applying artificial neural networks to a variety of tasks from fields such as computer vision and NLP. In this work we show that, by properly defining attention for convolutional neural networks, we can actually use this type of information in order to significantly improve the performance of a student CNN network by forcing it to mimic the attention maps of a powerful teacher network. To that end, we propose several novel methods of transferring attention, showing consistent improvement across a variety of datasets and convolutional neural network architectures. Code and models for our experiments are available at https://github.com/szagoruyko/attention-transfer

Over the last years, deep convolutional neural networks (ConvNets) have transformed the field of computer vision thanks to their unparalleled capacity to learn high level semantic image features. However, in order to successfully learn those features, they usually require massive amounts of manually labeled data, which is both expensive and impractical to scale. Therefore, unsupervised semantic feature learning, i.e., learning without requiring manual annotation effort, is of crucial importance in order to successfully harvest the vast amount of visual data that are available today. In our work we propose to learn image features by training ConvNets to recognize the 2d rotation that is applied to the image that it gets as input. We demonstrate both qualitatively and quantitatively that this apparently simple task actually provides a very powerful supervisory signal for semantic feature learning. We exhaustively evaluate our method in various unsupervised feature learning benchmarks and we exhibit in all of them state-of-the-art performance. Specifically, our results on those benchmarks demonstrate dramatic improvements w.r.t. prior state-of-the-art approaches in unsupervised representation learning and thus significantly close the gap with supervised feature learning. For instance, in PASCAL VOC 2007 detection task our unsupervised pre-trained AlexNet model achieves the state-of-the-art (among unsupervised methods) mAP of 54.4% that is only 2.4 points lower from the supervised case. We get similarly striking results when we transfer our unsupervised learned features on various other tasks, such as ImageNet classification, PASCAL classification, PASCAL segmentation, and CIFAR-10 classification. The code and models of our paper will be published on: https://github.com/gidariss/FeatureLearningRotNet .

In this paper we show how to learn directly from image data (i.e., without resorting to manually-designed features) a general similarity function for comparing image patches, which is a task of fundamental importance for many computer vision problems. To encode such a function, we opt for a CNN-based model that is trained to account for a wide variety of changes in image appearance. To that end, we explore and study multiple neural network architectures, which are specifically adapted to this task. We show that such an approach can significantly outperform the state-of-the-art on several problems and benchmark datasets.

IMGT/V-QUEST is the highly customized and integrated system for the standardized analysis of the immunoglobulin (IG) and T cell receptor (TR) rearranged nucleotide sequences. IMGT/V-QUEST identifies the variable (V), diversity (D) and joining (J) genes and alleles by alignment with the germline IG and TR gene and allele sequences of the IMGT reference directory. New functionalities were added through a complete rewrite in Java. IMGT/V-QUEST analyses batches of sequences (up to 50) in a single run. IMGT/V-QUEST describes the V-REGION mutations and identifies the hot spot positions in the closest germline V gene. IMGT/V-QUEST can detect insertions and deletions in the submitted sequences by reference to the IMGT unique numbering. IMGT/V-QUEST integrates IMGT/JunctionAnalysis for a detailed analysis of the V-J and V-D-J junctions, and IMGT/Automat for a full V-J- and V-D-J-REGION annotation. IMGT/V-QUEST displays, in 'Detailed view', the results and alignments for each submitted sequence individually and, in 'Synthesis view', the alignments of the sequences that, in a given run, express the same V gene and allele. The 'Advanced parameters' allow to modify default parameters used by IMGT/V-QUEST and IMGT/JunctionAnalysis according to the users' interest. IMGT/V-QUEST is freely available for academic research at http://imgt.cines.fr.

IMGT, the international ImMunoGeneTics information system (http://www.imgt.org), was created in 1989 by Marie-Paule Lefranc, Laboratoire d'ImmunoGénétique Moléculaire LIGM (Université Montpellier 2 and CNRS) at Montpellier, France, in order to standardize and manage the complexity of immunogenetics data. The building of a unique ontology, IMGT-ONTOLOGY, has made IMGT the global reference in immunogenetics and immunoinformatics. IMGT is a high-quality integrated knowledge resource specialized in the immunoglobulins or antibodies, T cell receptors, major histocompatibility complex, of human and other vertebrate species, proteins of the IgSF and MhcSF, and related proteins of the immune systems of any species. IMGT provides a common access to standardized data from genome, proteome, genetics and 3D structures. IMGT consists of five databases (IMGT/LIGM-DB, IMGT/GENE-DB, IMGT/3Dstructure-DB, etc.), fifteen interactive online tools for sequence, genome and 3D structure analysis, and more than 10,000 HTML pages of synthesis and knowledge. IMGT is used in medical research (autoimmune diseases, infectious diseases, AIDS, leukemias, lymphomas and myelomas), veterinary research, biotechnology related to antibody engineering (phage displays, combinatorial libraries, chimeric, humanized and human antibodies), diagnostics (clonalities, detection and follow-up of residual diseases) and therapeutical approaches (graft, immunotherapy, vaccinology). IMGT is freely available at http://www.imgt.org.

The human visual system has the remarkably ability to be able to effortlessly learn novel concepts from only a few examples. Mimicking the same behavior on machine learning vision systems is an interesting and very challenging research problem with many practical advantages on real world vision applications. In this context, the goal of our work is to devise a few-shot visual learning system that during test time it will be able to efficiently learn novel categories from only a few training data while at the same time it will not forget the initial categories on which it was trained (here called base categories). To achieve that goal we propose (a) to extend an object recognition system with an attention based few-shot classification weight generator, and (b) to redesign the classifier of a ConvNet model as the cosine similarity function between feature representations and classification weight vectors. The latter, apart from unifying the recognition of both novel and base categories, it also leads to feature representations that generalize better on "unseen" categories. We extensively evaluate our approach on Mini-ImageNet where we manage to improve the prior state-of-the-art on few-shot recognition (i.e., we achieve 56.20% and 73.00% on the 1-shot and 5-shot settings respectively) while at the same time we do not sacrifice any accuracy on the base categories, which is a characteristic that most prior approaches lack. Finally, we apply our approach on the recently introduced few-shot benchmark of Bharath and Girshick [4] where we also achieve state-of-the-art results.

We propose an object detection system that relies on a multi-region deep convolutional neural network (CNN) that also encodes semantic segmentation-aware features. The resulting CNN-based representation aims at capturing a diverse set of discriminative appearance factors and exhibits localization sensitivity that is essential for accurate object localization. We exploit the above properties of our recognition module by integrating it on an iterative localization mechanism that alternates between scoring a box proposal and refining its location with a deep CNN regression model. Thanks to the efficient use of our modules, we detect objects with very high localization accuracy. On the detection challenges of PASCAL VOC2007 and PASCAL VOC2012 we achieve mAP of 78.2% and 73.9% correspondingly, surpassing any other published work by a significant margin.

IMGT(®), the international ImMunoGeneTics information system(®)(http://www.imgt.org) is the global reference in immunogenetics and immunoinformatics. By its creation in 1989 by Marie-Paule Lefranc (Université de Montpellier and CNRS), IMGT(®) marked the advent of immunoinformatics, which emerged at the interface between immunogenetics and bioinformatics. IMGT(®) is specialized in the immunoglobulins (IG) or antibodies, T cell receptors (TR), major histocompatibility (MH) and proteins of the IgSF and MhSF superfamilies. IMGT(®) is built on the IMGT-ONTOLOGY axioms and concepts, which bridged the gap between genes, sequences and 3D structures. The concepts include the IMGT(®) standardized keywords (identification), IMGT(®) standardized labels (description), IMGT(®) standardized nomenclature (classification), IMGT unique numbering and IMGT Colliers de Perles (numerotation). IMGT(®) comprises 7 databases, 17 online tools and 15,000 pages of web resources, and provides a high-quality and integrated system for analysis of the genomic and expressed IG and TR repertoire of the adaptive immune responses, including NGS high-throughput data. Tools and databases are used in basic, veterinary and medical research, in clinical applications (mutation analysis in leukemia and lymphoma) and in antibody engineering and humanization. The IMGT/mAb-DB interface was developed for therapeutic antibodies and fusion proteins for immunological applications (FPIA). IMGT(®) is freely available at http://www.imgt.org.

IMGT/GENE-DB is the comprehensive IMGT genome database for immunoglobulin (IG) and T cell receptor (TR) genes from human and mouse, and, in development, from other vertebrates. IMGT/GENE-DB is the international reference for the IG and TR gene nomenclature and works in close collaboration with the HUGO Nomenclature Committee, Mouse Genome Database and genome committees for other species. IMGT/GENE-DB allows a search of IG and TR genes by locus, group and subgroup, which are CLASSIFICATION concepts of IMGT-ONTOLOGY. Short cuts allow the retrieval gene information by gene name or clone name. Direct links with configurable URL give access to information usable by humans or programs. An IMGT/GENE-DB entry displays accurate gene data related to genome (gene localization), allelic polymorphisms (number of alleles, IMGT reference sequences, functionality, etc.) gene expression (known cDNAs), proteins and structures (Protein displays, IMGT Colliers de Perles). It provides internal links to the IMGT sequence databases and to the IMGT Repertoire Web resources, and external links to genome and generalist sequence databases. IMGT/GENE-DB manages the IMGT reference directory used by the IMGT tools for IG and TR gene and allele comparison and assignment, and by the IMGT databases for gene data annotation. IMGT/GENE-DB is freely available at http://imgt.cines.fr.

This paper poses object category detection in images as a type of 2D-to-3D alignment problem, utilizing the large quantities of 3D CAD models that have been made publicly available online. Using the "chair" class as a running example, we propose an exemplar-based 3D category representation, which can explicitly model chairs of different styles as well as the large variation in viewpoint. We develop an approach to establish part-based correspondences between 3D CAD models and real photographs. This is achieved by (i) representing each 3D model using a set of view-dependent mid-level visual elements learned from synthesized views in a discriminative fashion, (ii) carefully calibrating the individual element detectors on a common dataset of negative images, and (iii) matching visual elements to the test image allowing for small mutual deformations but preserving the viewpoint and style constraints. We demonstrate the ability of our system to align 3D models with 2D objects in the challenging PASCAL VOC images, which depict a wide variety of chairs in complex scenes.

The Cypher property graph query language is an evolving language, originally designed and implemented as part of the Neo4j graph database, and it is currently used by several commercial database products and researchers. We describe Cypher 9, which is the first version of the language governed by the openCypher Implementers Group. We first introduce the language by example, and describe its uses in industry. We then provide a formal semantic definition of the core read-query features of Cypher, including its variant of the property graph data model, and its ASCII Art graph pattern matching mechanism for expressing subgraphs of interest to an application. We compare the features of Cypher to other property graph query languages, and describe extensions, at an advanced stage of development, which will form part of Cypher 10, turning the language into a compositional language which supports graph projections and multiple named graphs.

Under consideration is the large body of signal recovery problems that can be formulated as the problem of minimizing the sum of two (not necessarily smooth) lower semicontinuous convex functions in a real Hilbert space. This generic problem is analyzed and a decomposition method is proposed to solve it. The convergence of the method, which is based on the Douglas-Rachford algorithm for monotone operator-splitting, is obtained under general conditions. Applications to non-Gaussian image denoising in a tight frame are also demonstrated.

IMGT/V-QUEST is the highly customized and integrated online IMGT(®) tool for the standardized analysis of the immunoglobulin (IG) or antibody and T cell receptor (TR) rearranged nucleotide sequences. The analysis of these antigen receptors represents a crucial challenge for the study of the adaptive immune response in normal and disease-related situations. The expressed IG and TR repertoires represent a potential of 10(12) IG and 10(12) TR per individual. This huge diversity results from mechanisms that occur at the DNA level during the IG and TR molecular synthesis. These mechanisms include the combinatorial rearrangements of the variable (V), diversity (D) and joining (J) genes, the N-diversity (deletion and addition at random of nucleotides during the V-(D)-J rearrangement) and, for IG, somatic hypermutations. IMGT/V-QUEST identifies the V, D, J genes and alleles by alignment with the germline IG and TR gene and allele sequences of the IMGT reference directory. The tool describes the V-REGION mutations and identifies the hot spot positions in the closest germline V gene. IMGT/V-QUEST integrates IMGT/JunctionAnalysis for a detailed analysis of the V-J and V-D-J junctions and IMGT/Automat for a complete annotation of the sequences and also provides IMGT Collier de Perles. IMGT/HighV-QUEST, the high-throughput version of IMGT/V-QUEST, implemented to answer the needs of deep sequencing data analysis from Next Generation Sequencing (NGS), allows the analysis of thousands of IG and TR sequences in a single run. IMGT/V-QUEST and IMGT/HighV-QUEST are available at the IMGT(®) Home page, http://www.imgt.org.

Visual understanding is often based on measuring similarity between observations. Learning similarities specific to a certain perception task from a set of examples has been shown advantageous in various computer vision and pattern recognition problems. In many important applications, the data that one needs to compare come from different representations or modalities, and the similarity between such data operates on objects that may have different and often incommensurable structure and dimensionality. In this paper, we propose a framework for supervised similarity learning based on embedding the input data from two arbitrary spaces into the Hamming space. The mapping is expressed as a binary classification problem with positive and negative examples, and can be efficiently learned using boosting algorithms. The utility and efficiency of such a generic approach is demonstrated on several challenging applications including cross-representation shape retrieval and alignment of multi-modal medical images. 1.