HAL : Dernières publications

• [hal-03795954] Enumeration of Corner Polyhedra and 3-Connected Schnyder Labelings

  [...]

• [hal-03794815] A Model You Can Hear: Audio Identification with Playable Prototypes

  Machine learning techniques have proved useful for classifying and analyzing audio content. However, recent methods typically rely on abstract and high-dimensional representations that are difficult to interpret. Inspired by transformation-invariant approaches developed for image and 3D data, we propose an audio identification model based on learnable spectral prototypes. Equipped with dedicated transformation networks, these prototypes can be used to cluster and classify input audio samples from large collections of sounds. Our model can be trained with or without supervision and reaches state-of-the-art results for speaker and instrument identification, while remaining easily interpretable. The code is available at: https://github.com/romainloiseau/a-model-you-can-hear

• [hal-03794797] Online Segmentation of LiDAR Sequences: Dataset and Algorithm

  Roof-mounted spinning LiDAR sensors are widely used by autonomous vehicles. However, most semantic datasets and algorithms used for LiDAR sequence segmentation operate on $360^\circ$ frames, causing an acquisition latency incompatible with real-time applications. To address this issue, we first introduce HelixNet, a $10$ billion point dataset with fine-grained labels, timestamps, and sensor rotation information necessary to accurately assess the real-time readiness of segmentation algorithms. Second, we propose Helix4D, a compact and efficient spatio-temporal transformer architecture specifically designed for rotating LiDAR sequences. Helix4D operates on acquisition slices corresponding to a fraction of a full sensor rotation, significantly reducing the total latency. Helix4D reaches accuracy on par with the best segmentation algorithms on HelixNet and SemanticKITTI with a reduction of over $5\times$ in terms of latency and $50\times$ in model size. The code and data are available at: https://romainloiseau.fr/helixnet

• [hal-03793867] Share With Thy Neighbors: Single-View Reconstruction by Cross-Instance Consistency

  Approaches for single-view reconstruction typically rely on viewpoint annotations, silhouettes, the absence of background, multiple views of the same instance, a template shape, or symmetry. We avoid all such supervision and assumptions by explicitly leveraging the consistency between images of different object instances. As a result, our method can learn from large collections of unlabelled images depicting the same object category. Our main contributions are two ways for leveraging cross-instance consistency: (i) progressive conditioning, a training strategy to gradually specialize the model from category to instances in a curriculum learning fashion; and (ii) neighbor reconstruction, a loss enforcing consistency between instances having similar shape or texture. Also critical to the success of our method are: our structured autoencoding architecture decomposing an image into explicit shape, texture, pose, and background; an adapted formulation of differential rendering; and a new optimization scheme alternating between 3D and pose learning. We compare our approach, UNICORN, both on the diverse synthetic ShapeNet dataset - the classical benchmark for methods requiring multiple views as supervision - and on standard real-image benchmarks (Pascal3D+ Car, CUB) for which most methods require known templates and silhouette annotations. We also showcase applicability to more challenging real-world collections (CompCars, LSUN), where silhouettes are not available and images are not cropped around the object.

• [hal-03792466] Inline monitoring of 3D concrete printing using computer vision

  [...]

• [hal-03791973] Learning Uncertainty for Safety-Oriented Semantic Segmentation in Autonomous Driving

  [...]

• [hal-03791961] Latent reweighting, an almost free improvement for GANs

  Standard formulations of GANs, where a continuous function deforms a connected latent space, have been shown to be misspecified when fitting different classes of images. In particular, the generator will necessarily sample some lowquality images in between the classes. Rather than modifying the architecture, a line of works aims at improving the sampling quality from pre-trained generators at the expense of increased computational cost. Building on this, we introduce an additional network to predict latent importance weights and two associated sampling methods to avoid the poorest samples. This idea has several advantages: 1) it provides a way to inject disconnectedness into any GAN architecture, 2) since the rejection happens in the latent space, it avoids going through both the generator and the discriminator, saving computation time, 3) this importance weights formulation provides a principled way to reduce the Wasserstein's distance to the target distribution. We demonstrate the effectiveness of our method on several datasets, both synthetic and high-dimensional.

• [hal-03791786] TEACH: Temporal Action Composition for 3D Humans

  Given a series of natural language descriptions, our task is to generate 3D human motions that correspond semantically to the text, and follow the temporal order of the instructions. In particular, our goal is to enable the synthesis of a series of actions, which we refer to as temporal action composition. The current state of the art in text-conditioned motion synthesis only takes a single action or a single sentence as input. This is partially due to lack of suitable training data containing action sequences, but also due to the computational complexity of their non-autoregressive model formulation, which does not scale well to long sequences. In this work, we address both issues. First, we exploit the recent BABEL motion-text collection, which has a wide range of labeled actions, many of which occur in a sequence with transitions between them. Next, we design a Transformer-based approach that operates non-autoregressively within an action, but autoregressively within the sequence of actions. This hierarchical formulation proves effective in our experiments when compared with multiple baselines. Our approach, called TEACH for "TEmporal Action Compositions for Human motions", produces realistic human motions for a wide variety of actions and temporal compositions from language descriptions. To encourage work on this new task, we make our code available for research purposes at our website.

• [hal-03791776] TEMOS: Generating diverse human motions from textual descriptions

  We address the problem of generating diverse 3D human motions from textual descriptions. This challenging task requires joint modeling of both modalities: understanding and extracting useful human-centric information from the text, and then generating plausible and realistic sequences of human poses. In contrast to most previous work which focuses on generating a single, deterministic, motion from a textual description, we design a variational approach that can produce multiple diverse human motions. We propose TEMOS, a text-conditioned generative model leveraging variational autoencoder (VAE) training with human motion data, in combination with a text encoder that produces distribution parameters compatible with the VAE latent space. We show the TEMOS framework can produce both skeleton-based animations as in prior work, as well more expressive SMPL body motions. We evaluate our approach on the KIT Motion-Language benchmark and, despite being relatively straightforward, demonstrate significant improvements over the state of the art. Code and models are available on our webpage.

• [hal-03788879] Inline Monitoring of 3D Concrete Printing using Computer Vision

  The detection of anomalies is at the basis of any 3D printing control. In this paper, we propose a methodology for detection of anomalies based on computer vision. This methodology is composed of three modules: 1) image acquisition, 2) interlayer line and layer segmentation and 3) characterization of the local geometry and texture of the layers and detection of anomalies. The image acquisition is performed with a camera fixed to the printing nozzle. The proposed layer segmentation method recognizes and locates the lines separating the printed layers (F-score = 91%). The third module-taking as input the segmentation and the original image-evaluates the geometry of the layers and the texture of the material. The results are used to detect geometry anomalies when the values are outside the expected range. The material texture is classified into four classes of quality (macro-averaged F-score = 94%). We present the results and show the suitability of our methodology for automatic detection and localization of anomalies on images acquired during a printing session.

• [hal-03767854] The Evolution of the Idiolect over the Lifetime: A Quantitative and Qualitative Study of French 19th Century Literature

  The way in which authors express themselves is unique but changes over their lifetime. However, quantitative studies of this idiolectal evolution are rare. Using the Corpus for Idiolectal Research (CIDRE) that contains the dated works of 11 prolific 19th century French fiction writers, we propose new methods to identify, quantify and describe the grammatical-stylistic changes that take place using lexico-morphosyntactic patterns, also called motifs. To examine the strength of the chronological signal of change, we developed a method to calculate if a distance matrix of literary works contains a stronger chronological signal than expected by chance. Ten out of 11 corpora showed a higher than chance chronological signal, leading us to conclude that the evolution of the idiolect is in a mathematical sense monotonic, supporting the rectilinearity hypothesis previously put forward in the stylometric literature. The rectilinear property of the evolution of the idiolect found for most authors in CIDRE subsequently enabled us to propose a machine learning task: predicting the year in which a work was written. For the majority of the authors in our corpus, the accuracy and the amount of variance that is explained by the model were high and we discuss why the technique might fail for others. After applying a feature selection algorithm, we examined the most important features, i.e. the motifs that have the greatest influence on idiolectal evolution. We find that some of those features are stylistic and have been previously identified in qualitative literature studies. We report some remarkable stylistic constructions revealed by our algorithm to illustrate which kind of stylistic patterns can be extracted using our method.

• [tel-03783594] Efficient Approximations of High-Dimensional Data

  In this thesis, we study approximations of set systems (X,S), where X is a base set and S consists of subsets of X called ranges. Given a finite set system, our goal is to construct a small subset of X set such that each range is `well-approximated'. In particular, for a given parameter epsilon in (0,1), we say that a subset A of X is an epsilon-approximation of (X,S) if for any range R in S, the fractions |A cap R|/|A| and |R|/|X| are epsilon-close.Research on such approximations started in the 1950s, with random sampling being the key tool for showing their existence. Since then, the notion of approximations has become a fundamental structure across several communities---learning theory, statistics, combinatorics and algorithms. A breakthrough in the study of approximations dates back to 1971 when Vapnik and Chervonenkis studied set systems with finite VC-dimension, which turned out a key parameter to characterise their complexity. For instance, if a set system (X,S) has VC dimension d, then a uniform sample of O(d/epsilon^2) points is an epsilon-approximation of (X,S) with high probability. Importantly, the size of the approximation only depends on epsilon and d, and it is independent of the input sizes |X| and |S|!In the first part of this thesis, we give a modular, self-contained, intuitive proof of the above uniform sampling guarantee .In the second part, we give an improvement of a 30 year old algorithmic bottleneck---constructing matchings with low crossing number. This can be applied to build approximations with improved guarantees.Finally, we answer a 30 year old open problem of Blumer etal. by proving tight lower bounds on the VC dimension of unions of half-spaces - a geometric set system that appears in several applications, e.g. coreset constructions

• [hal-03329747] L’intensification subjective de noms intensifs et son corrélat : sur [un degré de N Modif]

  Nous décrivons les distributions syntaxiques du nom degré en combinaison avec des noms abstraits (qualité, état, procès), notamment dans les SN indéfinis. La sémantique des grandeurs intensives (Van de Velde 1995) dénotées par des noms abstraits biaise la référence de tels SN, pour privilégier la zone supérieure d’une grandeur intensive. À cause de la nature évidemment subjective d’une telle catégorisation, le locuteur ne se contente généralement pas de donner une telle prédication, mais l’accompagne souvent en donnant le corrélat du degré atteint, lorsque le discours porte exactement sur cet aspect. Pour ce faire, il mobilise divers moyens syntaxiques allant de la parataxe à l’hypotaxe.

• [hal-03780268] Ouverture

  [...]

• [hal-03776235] ZeBraInspector, a platform for the automated segmentation and analysis of body and brain volumes in whole 5 days post-fertilization zebrafish following simultaneous visualization with identical orientations

  In recent years, the zebrafish has become a well-established laboratory model. We describe here the ZeBraInspector (ZBI) platform for high-content 3D imaging (HCI) of 5 days post-fertilization zebrafish eleuthero-embryos (EEs). This platform includes a mounting method based on 3D-printed stamps to create a grid of wells in an agarose cast, facilitating batch acquisitions with a fast-confocal laser scanning microscope. We describe reference labeling in cleared fish with a fluorescent lipophilic dye. Based on this labeling, the ZBI software registers. EE 3D images, making it possible to visualize numerous identically oriented EEs on a single screen, and to compare their morphologies and any fluorescent patterns at a glance. High-resolution 2D snapshots can be extracted. ZBI software is therefore useful for diverse high-content analyses (HCAs). Following automated segmentation of the lipophilic dye signal, the ZBI software performs volumetric analyses on whole EEs and their nervous system white matter. Through two examples, we illustrate the power of these analyses for obtaining statistically significant results from a small number of samples: the characterization of a phenotype associated with a neurodevelopmental mutation, and of the defects caused by treatments with a toxic anti-cancer compound.

• [hal-03699921] Global Solution of Quadratic Problems by Interval Methods and Convex Reformulation

  [...]

• [hal-03360817] Electric Bus Charging Scheduling Strategy with Stochastic Arrival Time and State of Charge

  To alleviate the range anxiety of drivers and time-consuming charging for electric buses (eBuses), opportunity fast-charging has gradually been utilized. Considering that eBuses have operational tasks, identifying an optimal charging scheduling will be needed. However, in the real world, arrival time and state of charge (SOC) of eBuses are uncertain. Therefore, it is challenging for the charging station to efficiently schedule charging tasks. To solve the problem, this paper develops a two-stage stochastic eBus charging scheduling model. In the first stage, eBuses are assigned to designated chargers. After the arrival time and SOC are realized, the second stage determines the charging sequence of eBuses on each charger. The objective is to minimize the penalty cost of tardiness by determining the charging start time and the corresponding charging duration time. Then, a sample average approximation (SAA) algorithm is applied. Additional numerical experiments are performed to verify the efficiency of the stochastic programming model and algorithm.

• [hal-03765422] TokenCut: Segmenting Objects in Images and Videos with Self-supervised Transformer and Normalized Cut

  In this paper, we describe a graph-based algorithm that uses the features obtained by a self-supervised transformer to detect and segment salient objects in images and videos. With this approach, the image patches that compose an image or video are organised into a fully connected graph, where the edge between each pair of patches is labeled with a similarity score between patches using features learned by the transformer. Detection and segmentation of salient objects is then formulated as a graph-cut problem and solved using the classical Normalized Cut algorithm. Despite the simplicity of this approach, it achieves state-of-the-art results on several common image and video detection and segmentation tasks. For unsupervised object discovery, this approach outperforms the competing approaches by a margin of 6.1%, 5.7%, and 2.6%, respectively, when tested with the VOC07, VOC12, and COCO20K datasets. For the unsupervised saliency detection task in images, this method improves the score for Intersection over Union (IoU) by 4.4%, 5.6% and 5.2%. When tested with the ECSSD, DUTS, and DUT-OMRON datasets, respectively, compared to current state-of-the-art techniques. This method also achieves competitive results for unsupervised video object segmentation tasks with the DAVIS, SegTV2, and FBMS datasets.

• [hal-03666488] A bound for Delaunay flip algorithms on flat tori

  We are interested in triangulations of flat tori. A Delaunay flip algorithm performs Delaunay flips on the edges of an input triangulation T until it reaches a Delaunay triangulation. We prove that no sequence of Delaunay flips is longer than C_Γ • n^2 • Λ(T) where Λ(T) is the maximum length of an edge of T , n is the number of vertices of T , and C_Γ > 0 depends only on the flat torus. The bound improves on the preexisting upper bound in three ways: the dependency in the "quality" of the input triangulation is linear instead of quadratic, the bound is tight, and the "quality parameter" is simpler.

• [hal-03771031] Revisiting Semiring Provenance for Datalog

  Data provenance consists in bookkeeping meta information during query evaluation, in order to enrich query results with their trust level, likelihood, evaluation cost, and more. The framework of semiring provenance abstracts from the specific kind of meta information that annotates the data. While the definition of semiring provenance is uncontroversial for unions of conjunctive queries, the picture is less clear for Datalog. Indeed, the original definition might include infinite computations, and is not consistent with other proposals for Datalog semantics over annotated data. In this work, we propose and investigate several provenance semantics, based on different approaches for defining classical Datalog semantics. We study the relationship between these semantics, and introduce properties that allow us to analyze and compare them.

• [hal-03765971] Trust-based gateway selection in a multi-tiers blockchain architecture

  [...]

• [hal-03765961] Tactile Internet: Technologies, test platforms, trials, and applications

  [...]

• [hal-03765951] Toward a Machine Learning and Software Defined Network Approaches to Manage Miners’ Reputation in Blockchain

  [...]

• [hal-03765933] Taxonomy of Machine learning techniques and elastic resource optimization in sliced networks

  [...]

• [hal-03765925] Trust-based gateway selection in a multi-tiers blockchain architecture

  [...]

• [hal-03765919] IoT efficient data exploitation process using Model Driven Engineering

  [...]

• [hal-03765915] RoofCoin: a blockchain for internet of vehicles based on the ROOF standard and VSN

  [...]

• [hal-03765898] Computing offloading and load balancing within UAV clusters

  [...]

• [hal-03765221] Sorting Genomes by Prefix Double-Cut-and-Joins

  In this paper, we study the problem of sorting unichromosomal linear genomes by prefix double-cut-and-joins (or DCJs) in both the signed and the unsigned settings. Prefix DCJs cut the leftmost segment of a genome and any other segment, and recombine the severed endpoints in one of two possible ways: one of these options corresponds to a prefix reversal, which reverses the order of elements between the two cuts (as well as their signs in the signed case). Depending on whether we consider both options or reversals only, our main results are: (1) new structural lower bounds based on the breakpoint graph for sorting by unsigned prefix reversals, unsigned prefix DCJs, or signed prefix DCJs; (2) a polynomial-time algorithm for sorting by signed prefix DCJs, thus answering an open question in [8]; (3) a 3/2-approximation for sorting by unsigned prefix DCJs, which is, to the best of our knowledge, the first sorting by {\em prefix} rearrangements problem that admits an approximation ratio strictly smaller than 2 (with the obvious exception of the polynomial-time solvable problems); and finally, (4) an FPT algorithm for sorting by unsigned prefix DCJs parameterised by the number of breakpoints in the genome.

• [tel-03764907] Registration of Heterogenous Data for Urban Modeling

  This thesis is part of the Building Indoor/Outdoor Modeling (BIOM) project that aims at automatic, simultaneous indoor and outdoor modelling of buildings from heterogeneous data. The heterogeneity is both in data type (image and Light Detection and Ranging (LiDAR)) and acquisition platform: terrestrial indoor/outdoor or aerial acquisition. The first issue of such modeling is thus to precisely register this data. The work carried out has confirmed that the environment and the type of data drive the choice of the registration algorithm. Our contribution consists in exploring the fundamental properties of the data and the acquisition platforms in order to propose potential solutions for all the registration problems encountered by the project. As in abuilding environment, most objects are composed of geometric primitives (planarpolygons, straight lines, openings), we chose to introduce registration algorithmsbased on these primitives. The basic idea of these algorithms consists in the definition of a global energy between the extracted primitives from the data-sets to register and the proposal of a robust method for optimizing this energy based on the RANSAC paradigm. Our contribution ranging from the proposal of robust methods to extract the selected primitives to the integration of these primitives in an efficient registration framework. Our solutions have exceeded the limitations of existing algorithms and have proven their effectiveness in solving the challenging problems encountered by the project such as the indoor/outdoor registration, image/LiDAR registration, and aerial/terrestrialregistration.

• [hal-03132172] Compacted binary trees admit a stretched exponential

  [...]

• [hal-00808005] ECAR: an Energy/Channel Aware Routing Protocol For Cooperative Wireless Sensor Networks

  The proliferation of low power networks like Wireless Sensor Networks (WSN) rose up new challenges. Power conservation and channel quality become the most important parameters. Obviously, hop count based routing protocols are no more adapted to such networks having power limitations and channel problems. Several alternatives were suggested to cope with these constraints. In MAC layer for example, cooperative protocols were designed to enhance the channel use: the neighbor nodes help the source to retransmit its packets. However, if the path proposed by the routing protocol contains poor channels, the cooperative communications will not save all the packets. Therefore, the design of new routing protocol becomes compulsory. In this paper we propose ECAR, a routing protocol that optimizes two objectives at the same time: energy and Channel State Information (CSI). Compared to AODV, ECAR provides considerable enhancements in delivery ratio, end-to-end delay and power consumption.

• [tel-03760082] Diffeomorphisms, Optimal Transport and Applications to Imaging

  [...]

• [hal-03759300] Identifier, représenter, penser les réseaux urbains de l’humanisme autour de Marguerite de Valois : l’exemple du voyage des Flandres (1577)

  [...]

• [hal-00808020] RBCR: a Relay-Based Cooperative Routing Protocol for Cooperative Wireless Sensor Network

  In wireless sensor networks several constraints decrease communications performances. The main objective of this paper is to present a multi-objective routing algorithm RBCR that computes routing path based on the energy consumption and channel qualities. Additionally, the channel qualities are evaluated based on the presence of relay nodes. Compared to AODV and AODV associated to a cooperative MAC protocol, RBCR provides better performances.

• [hal-00808012] A Cooperative Low Power Mac Protocol for Wireless Sensor Networks

  Over the last decade cooperative communication in wireless sensor networks (WSN) received much attention. A lot of works have been done to propose a MAC layer that supports cooperative communication. However the impact of the association of a cooperative communication technique with a low power listening scheme was not studied in the literature. In this paper we propose CL-MAC, a Cooperative Low power mac protocol for WSNs. CL-MAC implements jointly Low Power Listening and cooperative communication. More precisely, we propose two variants of this protocol: a proactive version CL-MAC(P) and a reactive version CL-MAC(R). In order to evaluate the performances of the two proposed CL-MAC variants, we compare its to those of X-MAC. Simulation results proved that our protocol is able to enhance the use of the channel and to reach promising energy preservation especially in dense networks.

• [hal-03492186] Noncommutative unicellular LLT polynomials

  [...]

• [hal-03015531] Hardness of Minimum Barrier Shrinkage and Minimum Installation Path

  [...]

• [hal-02510473] DIABLO: Dictionary-based attention block for deep metric learning

  Recent breakthroughs in representation learning of unseen classes and examples have been made in deep metric learning by training at the same time the image representations and a corresponding metric with deep networks. Recent contributions mostly address the training part (loss functions, sampling strategies, etc.), while a few works focus on improving the discriminative power of the image representation. In this paper, we propose DIABLO, a dictionary-based attention method for image embedding. DIABLO produces richer representations by aggregating only visually-related features together while being easier to train than other attention-based methods in deep metric learning. This is experimentally confirmed on four deep metric learning datasets (Cub-200-2011, Cars-196, Stanford Online Products, and In-Shop Clothes Retrieval) for which DIABLO shows state-of-the-art performances.

• [hal-01289784] Hopf algebras of m-permutations, (m + 1)-ary trees, and m-parking functions

  [...]

• [hal-03676377] Automated design of dynamic programming schemes for RNA folding with pseudoknots

  Despite being a textbook application of dynamic programming (DP) and routine task in RNA structure analysis, RNA secondary structure prediction remains challenging whenever pseudoknots come into play. To circumvent the NP-hardness of energy minimization in realistic energy models, specialized algorithms have been proposed for restricted conformation classes that capture the most frequently observed configurations. While these methods rely on hand-crafted DP schemes, we generalize and fully automatize the design of DP pseudoknot prediction algorithms. We formalize the problem of designing DP algorithms for an (infinite) class of conformations, modeled by (a finite number of) fatgraphs, and automatically build DP schemes minimizing their algorithmic complexity. We propose an algorithm for the problem, based on the tree-decomposition of a well-chosen representative structure, which we simplify and reinterpret as a DP scheme. The algorithm is fixed-parameter tractable for the tree-width tw of the fatgraph, and its output represents a O(n tw+1) algorithm for predicting the MFE folding of an RNA of length n. Our general framework supports general energy models, partition function computations, recursive substructures and partial folding, and could pave the way for algebraic dynamic programming beyond the context-free case.

• [hal-03746074] Dimension Groups and Dynamical Systems

  We give a description of the link between topological dynamical systems and their dimension groups. The focus is on minimal systems and, in particular, on substitution shifts. We describe in detail the various classes of systems including Sturmian shifts and interval exchange shifts. This is a preliminary version of a book which will be published by Cambridge University Press. Any comments are of course welcome.

• [hal-03745835] Coverage Extension of Software Defined Radio Platforms for 3GPP 4G/5G Radio Access Networks

  The quick development of Software Defined Radio (SDR) platforms has simplified experimenting 3GPP cellular wireless access technologies in software. Thus, many implementations of 3GPP 4G/5G SDR Radio Access Networks (S-RANs) have been developed in the literature. However, most of these S-RANs are tested and evaluated in laboratories within very short-range cells. One of the main reasons for this restricted test environment is the limitation of the output transmit power of SDR devices. The objective of this paper is to show how the cell coverage of 3GPP 4G/5G S-RANs can be extended (numerically and experimentally) to reach a desired cell size such as micro and macro-cells. For this purpose, we first model the communication between user equipment (UE) and eNodeB/gNodeB using an SDR-based radio transceiver in both uplink and downlink directions. Then, we extract the parameters affecting the transmitted and received signal power. After, we numerically analyze the impact of each parameter on the cell coverage. Finally, to validate the numerical analysis, we conduct experiments using a testbed based on an open-source S-RAN implementation to evaluate the transmit power, receive power, Signal-to-Noise Ratio (SNR), and Block Error Rate (BLER). The obtained numerical and experimental results show that different 4G/5G cell sizes can be achieved by using appropriate external equipment including amplifiers, filters, and antennas calibrated according to the employed SDR device specification.

• [hal-03740481] Improving Deep Metric Learning With Virtual Classes And Examples Mining

  In deep metric learning, the training procedure relies on sampling informative tuples. However, as the training procedure progresses, it becomes nearly impossible to sample relevant hard negative examples without proper mining strategies or generation-based methods. Recent work on hard negative generation have shown great promises to solve the mining problem. However, this generation process is difficult to tune and often leads to incorrectly labeled examples. To tackle this issue, we introduce MIRAGE, a generation-based method that relies on virtual classes entirely composed of generated examples that act as buffer areas between the training classes. We empirically show that virtual classes significantly improve the results on popular datasets (Cub-200-2011 and Cars-196) compared to other generation methods

• [hal-01418393] Dual Block Coordinate Forward-Backward Algorithm with Application to Deconvolution and Deinterlacing of Video Sequences

  Optimization methods play a central role in the solution of a wide array of problems encountered in various application fields, such as signal and image processing. Especially when the problems are highly dimensional, proximal methods have shown their efficiency through their capability to deal with composite, possibly nonsmooth objective functions. The cornerstone of these approaches is the proximity operator, which has become a quite popular tool in optimization. In this work, we propose new dual forward-backward formulations for computing the proximity operator of a sum of convex functions involving linear operators. The proposed algorithms are accelerated thanks to the introduction of a block-coordinate strategy combined with a preconditioning technique. Numerical simulations emphasize the good performance of our approach for the problem of jointly deconvoluting and deinterlacing video sequences.

• [hal-03737125] Online convex optimization in wireless networks and beyond: The feedback -performance trade-off

  The high degree of variability present in current and emerging mobile wireless networks calls for mathematical tools and techniques that transcend classical (convex) optimization paradigms. The aim of this short survey paper is to provide a gentle introduction to online learning and optimization algorithms that are able to provably cope with this variability and provide policies that are asymptotically optimal in hindsight-a property known as no regret. The focal point of this survey will be to delineate the trade-off between the information available as feedback to the learner, and the achievable regret guaranteesstarting with the case of gradient-based (first-order) feedback, then moving on to value-based (zeroth-order) feedback, and, ultimately, pushing the envelope to the extreme case of a single bit of feedback. We illustrate our theoretical analysis with a series of practical wireless network examples that highlight the potential of this elegant toolbox.

• [hal-02453173] The Steep-Bounce zeta map in Parabolic Cataland

  [...]

• [hal-02314445] Human pose regression by combining indirect part detection and contextual information

  In this paper, we tackle the problem of human pose estimation from still images, which is a very active topic, specially due to its several applications, from image annotation to human-machine interface. We use the soft-argmax function to convert feature maps directly to body joint coordinates, resulting in a fully differentiable framework. Our method is able to learn heat maps representations indirectly, without additional steps of artificial ground truth generation. Consequently, contextual information can be included to the pose predictions in a seamless way. We evaluated our method on two challenging datasets, the Leeds Sports Poses (LSP) and the MPII Human Pose datasets, reaching the best performance among all the existing regression methods. Source code available at: https://github.com/dluvizon/pose-regression.

• [hal-02414845] Optimal bounds for computing α-gapped repeats

  [...]

• [hal-01978633] Exploring Weight Symmetry in Deep Neural Networks

  We propose to impose symmetry in neural network parameters to improve parameter usage and make use of dedicated convolution and matrix multiplication routines. Due to significant reduction in the number of parameters as a result of the symmetry constraints, one would expect a dramatic drop in accuracy. Surprisingly, we show that this is not the case, and, depending on network size, symmetry can have little or no negative effect on network accuracy, especially in deep overparameterized networks. We propose several ways to impose local symmetry in recurrent and convolutional neural networks, and show that our symmetry parameterizations satisfy universal approximation property for single hidden layer networks. We extensively evaluate these parameterizations on CIFAR, ImageNet and language modeling datasets, showing significant benefits from the use of symmetry. For instance, our ResNet-101 with channel-wise symmetry has almost 25% less parameters and only 0.2% accuracy loss on ImageNet. Code for our experiments is available at https://github.com/hushell/deep-symmetry