Search

International audience; Rising greenhouse gas emissions are causing increasing worldwide impacts and changes on climate patterns, sea level, food production, human lives and livelihoods. Maintaining or improving the ability of coastal aquatic ecosystems and oceans to remove CO2 from the atmosphere is a crucial aspect for climate mitigation. The vegetated coastal ecosystems are able to catch and to sequestrate carbon, the so-called Blue Carbon. These ecosystems are key exchange zones that mediate the biogeochemical cycles across the continent, the ocean and the atmosphere. Given the importance of these ecosystems in biogeochemical cycles and their sensitivity to natural and anthropogenic pressures, the carbon cycle within and between compartments (e.g. pelagos, benthos…) and across the interfaces (e.g. atmosphere, ocean…) need to be addressed. On a regional scale, in the extended urban area of La Rochelle located on the French Atlantic coast (La Rochelle metropolitan area), lack in situ measurements within the wetlands and littoral zone make very uncertain their role as a sink or a source of CO2 to the atmosphere. We will first present a vast research project “La Rochelle Territoire Zéro Carbone” project (https://www.agglo-larochelle.fr/projet-de-territoire/territoire-zero-carbone), that target the ambition of carbon neutrality in La Rochelle metropolitan area by 2040 through a holistic approach (from measuring CO2 to raising people's awareness and assessing the impact of exogenous natural factors). Second, we will present some first results on the Blue Carbon dynamics within the freshwater and salt marshes, and seagrasses.

Moving Object Detection (MOD) is a fundamental step for many computer vision applications. MOD becomes very challenging when a video sequence captured from a static or moving camera suffers from the challenges: camouflage, shadow, dynamic backgrounds, and lighting variations, to name a few. Deep learning methods have been successfully applied to address MOD with competitive performance. However, in order to handle the overfitting problem, deep learning methods require a large amount of labeled data which is a laborious task as exhaustive annotations are always not available. Moreover, some MOD deep learning methods show performance degradation in the presence of unseen video sequences because the testing and training splits of the same sequences are involved during the network learning process. In this work, we pose the problem of MOD as a node classification problem using Graph Convolutional Neural Networks (GCNNs). Our algorithm, dubbed as GraphMOD-Net, encompasses instance segmentation, background initialization, feature extraction, and graph construction. GraphMOD-Net is tested on unseen videos and outperforms state-of-the-art methods in unsupervised, semi-supervised, and supervised learning in several challenges of the Change Detection 2014 (CDNet2014) and UCSD background subtraction datasets.

International audience; We present a novel microscopic tridomain model describing the electrical activity in cardiac tissue with dynamical gap junctions. The microscopic tridomain system consists of three PDEs modeling the tissue electrical conduction in the intra-and extra-cellular domains, supplemented by a nonlinear ODE system for the dynamics of the ion channels and the gap junctions. We establish the global existence and uniqueness of the weak solutions to our microscopic tridomain model. The global existence of solution, which constitutes the main result of this paper, is proved by means of an approximate non-degenerate system, the Faedo-Galerkin method, and an appropriate compactness argument.

In this paper, we are dealing with a rigorous homogenization result at two different levels for the bidomain model of cardiac electro-physiology. The first level associated with the mesoscopic structure such that the cardiac tissue consists of extracellular and intracellular domains separated by an interface (the sarcolemma). The second one related to the microscopic structure in such a way that the intracellular medium can only be viewed as a periodical layout of unit cells (mitochondria). At the interface between intra-and extracellular media, the fluxes are given by nonlinear functions of ionic and applied currents. A rigorous homogenization process based on unfolding operators is applied to derive the macroscopic (homogenized) model from our meso-microscopic bidomain model. We apply a three-scale unfolding method in the intracellular problem to obtain its homogenized equation at two levels. The first level upscaling of the intracellular structure yields the mesoscopic equation. The second step of the homogenization leads to obtain the intracellular homogenized equation. To prove the convergence of the nonlinear terms, especially those defined on the microscopic interface, we use the boundary unfolding method and a Kolmogorov-Riesz compactness's result. Next, we use the standard unfolding method to homogenize the extracellular problem. Finally, we obtain, at the limit, a reaction-diffusion system on a single domain (the superposition of the intracellular and extracellular media) which contains the homogenized equations depending on three scales. Such a model is widely used for describing the macroscopic behavior of the cardiac tissue, which is recognized to be an important messengers between the cytoplasm (intracellular) and the other extracellular inside the biological cells. Contents 2020 Mathematics Subject Classification. 65N55 and 35A01 and 35B27 and 35K57 and 65M.., arXiv admin note: substantial text overlap with arXiv:2201.03412

The paper is devoted to the mathematical analysis of the Cauchy problem for general cross-diffusion systems without any assumption about its entropic structure. A global existence result of nonnegative solutions is obtained by applying a classical Schauder fixed point theorem. The proof is upgraded for enhancing the regularity of the solution, namely its gradient belongs to the space L r ( ( 0 , T ) × Ω ) for some r > 2 . To this aim, the Schauder's strategy is coupled with an extension of Meyers regularity result for linear parabolic equations. We show how this approach allows to prove the well-posedness of the problem using only assumptions prescribing and admissibility range for the ratios between the diffusion and cross-diffusion coefficients. The results are compared with those that are reachable with an additional regularity assumption on the parabolic operator, namely a small BMO assumption for its coefficients. Finally, the question of the maximal principle is also addressed, especially when source terms are incorporated in the equation in order to ensure the confinement of the solution.

International audience; In the present paper, a new three-scale asymptotic homogenization method is proposed to study the electrical behavior of the cardiac tissue structure with multiple heterogeneities at two different levels. The first level is associated with the mesoscopic structure such that the cardiac tissue is composed of extracellular and intracellular domains. The second level is associated with the microscopic structure in such a way the intracellular medium can only be viewed as a periodical layout of unit cells (mitochondria). Then, we define two kinds of local cells that are obtained by upscaling methods. The homogenization method is based on a power series expansion which allows determining the macroscopic (homogenized) bidomain model from the microscopic bidomain problem at each structural level. First, we use the two-scale asymptotic expansion to homogenize the extracellular problem. Then, we apply a three-scale asymptotic expansion in the intracellular problem to obtain its homogenized equation at two levels. The first upscaling level of the intracellular structure yields the mesoscopic equation and the second step of the homogenization leads to obtain the intracellular homogenized equation. Both the mesoscopic and microscopic information is obtained by homogenization to capture local characteristics inside the cardiac tissue structure. Finally, we obtain the macroscopic bidomain model and the heart domain coincides with the intracellular medium and extracellular one, which are two inter-penetrating and superimposed continua connected at each point by the cardiac cellular membrane. The interest of the proposed method comes from the fact that it combines microscopic and mesoscopic characteristics to obtain a macroscopic description of the electrical behavior of the heart.

International audience

International audience

This paper proposes a fusion method of modalities extracted from video through a three-stream network with spatio-temporal and temporal convolutions for fine-grained action classification in sport. It is applied to TTStroke-21 dataset which consists of untrimmed videos of table tennis games. The goal is to detect and classify table tennis strokes in the videos, the first step of a bigger scheme aiming at giving feedback to the players for improving their performance. The three modalities are raw RGB data, the computed optical flow and the estimated pose of the player. The network consists of three branches with attention blocks. Features are fused at the latest stage of the network using bilinear layers. Compared to previous approaches, the use of three modalities allows faster convergence and better performances on both tasks: classification of strokes with known temporal boundaries and joint segmentation and classification. The pose is also further investigated in order to offer richer feedback to the athletes., Comment: MMSports '21, October 20, 2021, Virtual Event,, Oct 2021, Chengdu, China

International audience; Les couches de type Dense et Conv2d constituent parmi les briques élémentaires les plus populaires des réseaux de neurones actuels. Dans cet article, nous proposons d’en explorer une variante où les hyperplans sont remplacés par des hypersphères. Nous utilisons pour cela le modèle conforme défini par Hestenes et al. [6] où des hypersphères en dimension n peuvent être paramétrées par un vecteur de l’algèbre géométrique conforme Rn+1,1.En ce qui concerne la couche Dense, une approche similaire existe déjà dans Banarer et al. [1]. L’implémentation de ces deux nouveaux types de couche via des opérateurs d’algèbre linéaire/de convolution classiques est explicitée en même temps que les contraintes pour la mise à jour des poids. L’observation d’expérimentations réalisées sur des données synthétiques et des bases d’images conduit à conclure que les variantes « sphériques » des couches Dense et Conv2d améliorent la fonction de coût et convergent plus rapidement avec un nombre identiques de paramètres. Malheureusement, ce type de couches se révèlent parfois plus sensibles à l’initialisation ce qui peut conduire à une certaine instabilité. Quelques pistes pour remédier à ces problèmes sont présentées.

The paper entitled "Well posedness of general cross-diffusion systems", by C. Choquet, C. Rosier, L. Rosier, J. Diff. Eq. 2021, is devoted to the mathematical analysis of the Cauchy problem for general cross-diffusion systems without any assumption about its entropic structure. The absence of this type of hypothesis is strongly felt for two questions: the uniqueness of the solution, despite the nonlinear coupling of the highest derivatives terms, and the maximum principle. The article is therefore largely devoted to these two points. The answers are provided at the cost of certain assumptions or technicalities, mainly: (i) the ratios between the diffusion and cross-diffusion coefficients has to be drastically controlled for sufficiently enhancing the regularity of the solution, namely its gradient belongs to the space $ L ^ 4 ((0,T)\times \Omega )$; the regularity is obtained by adapting the classical Meyer's to the nonlinear parabolic setting under consideration ; (ii) the source terms have to ensure the confinement of the solution. The present "companion" paper aims at showing where more classical analysis tools fail to solve these questions and gives some additional clarifications., Comment: 16 pages, 2 figures

A new numerical method for the identification of initial conditions in wave propagation or heat conduction problems is constructed. The method is based on the wavelet basis Galerkin discretization of the model equations, without time discretization as in classical approches. A singular value decomposition is applied to the resulting discrete system to get explicitly its solution, and then to reconstruct the initial condition from measurements. Numerical results on toy solutions show the efficiency and the accuracy of the method.

CBMI 2018 was organized in cooperation with the IEEE Signal Processing Society and the European Association for Signal Processing (EURASIP); International audience

International audience

International audience; This essay inquires how mathematical beings could be inserted into the architecture of modes of existence proposed by Bruno Latour in the framework of his pluralist and renewed ontology of the modern world. An answer to this question is put forward with the aid of the work of Reviel Netz on the emergence of Greek mathematics, and of Charles Sanders Peirce on the diagrammatic dimension of mathematical practice, in the framework of an empiric conception of mathematics based on the notion of experience according to William James, and inspired by certain aspects of Per Martin-Löf's philosophy. It provides a way of describing the firm certainty with which proofs endow theorems, while invalidating the interpretation of this certainty as the mark of a direct access to an absolute and transcendental truth.; L'objet de cet essai est l'accueil des entités mathématiques dans l'architecture des modes d'existence proposée par Bruno Latour dans le cadre de son ontologie pluraliste du monde moderne. Les travaux de Reviel Netz sur l'émergence des mathématiques grecques et de Charles Sanders Peirce sur la dimension diagrammatique de l'activité mathématique sont employés pour proposer une réponse dans le cadre d'une conception empirique des mathématiques basée sur la notion d'expérience chère à William James et inspirée par certains aspects de la philosophie de Per Martin-Löf. Cette approche permet de penser la solide certitude dont la démonstration dote les résultats mathématiques tout en invalidant son interprétation comme la marque d'un accès direct à une vérité absolue et transcendante.

Conventional neural networks show a powerful framework for background subtraction in video acquired by static cameras. Indeed, the well-known SOBS method and its variants based on neural networks were the leader methods on the largescale CDnet 2012 dataset during a long time. Recently, convolutional neural networks which belong to deep learning methods were employed with success for background initialization, foreground detection and deep learned features. Currently, the top current background subtraction methods in CDnet 2014 are based on deep neural networks with a large gap of performance in comparison on the conventional unsupervised approaches based on multi-features or multi-cues strategies. Furthermore, a huge amount of papers was published since 2016 when Braham and Van Droogenbroeck published their first work on CNN applied to background subtraction providing a regular gain of performance. In this context, we provide the first review of deep neural network concepts in background subtraction for novices and experts in order to analyze this success and to provide further directions. For this, we first surveyed the methods used background initialization, background subtraction and deep learned features. Then, we discuss the adequacy of deep neural networks for background subtraction. Finally, experimental results are presented on the CDnet 2014 dataset., Comment: 46 pages, 4 figures, submitted to neural networks

Foreground segmentation is a fundamental problem in many artificial intelligence and computer vision based applications. However, robust foreground segmentation with high precision is still a challenging problem in complex scenes. Currently, many of the existing algorithms process the input data in RGB space only, where the foreground segmentation performance is most likely degraded by various challenges like shadows, color camouflage, illumination changes, out of range camera sensors and bootstrapping. Cameras capturing RGBD data are highly active visual sensors as they provide depth information along with RGB of the given input images. Therefore, to address the challenging problem we propose a foreground segmentation algorithm based on conditional generative adversarial networks using RGB and depth data. The goal of our proposed model is to perform robust foreground segmentation in the presence of various complex scenes with high accuracy. For this purpose, we trained our GAN based CNN model with RGBD input data conditioned on ground-truth information in an adversarial fashion. During training, our proposed model aims to learn the foreground segmentation on the basis of cross-entropy loss and euclidean distance loss to identify between real vs fake samples. While during testing the model is given RGBD input to the trained generator network that performs robust foreground segmentation. Our proposed method is evaluated using two RGBD benchmark datasets that are SBM-RGBD and MULTIVISION kinect. Various experimental evaluations and comparative analysis of our proposed model with eleven existing methods confirm its superior performance.

Moving Object Segmentation (MOS) is an important topic in computer vision. MOS becomes a challenging problem in the presence of dynamic background and moving camera videos such as Pan–Tilt–Zoom cameras (PTZ). The MOS problem has been solved using unsupervised and supervised learning strategies. Recently, new ideas to solve MOS using semi-supervised learning have emerged inspired from the theory of Graph Signal Processing (GSP). These new algorithms are usually composed of several steps including: segmentation, background initialization, features extraction, graph construction, graph signal sampling, and a semi-supervised learning algorithm inspired from reconstruction of graph signals. In this work, we summarize and explain the theoretical foundations as well as the technical details of MOS using GPS. We also propose two architectures for MOS using semi-supervised learning and a new evaluation procedure for GSP-based MOS algorithms. GSP-based algorithms are evaluated in the Change Detection (CDNet2014) dataset for MOS, outperforming numerous State-Of-The-Art (SOTA) methods in several challenging conditions.

International audience; This article aims at providing an efficient algorithm for sea clutter segmentation in coherent radar images. Using multiphase interface motion models, we show that segmentation in the range–time–intensity map gives much more relevant results than segmentation in the range–Doppler Map and avoids the use of assumptions on the clutter frequency response. Several mathematical models are evaluated on real data corresponding to Douglas sea states of approximately 5. Dealing with diffuse interface motion models involving more than two phases allows us to take into account the low contrast, the interference noise, and the contour weakness that are typical of range–time–intensity map images. Sea clutter segmentation allows us, for instance, to obtain oceanographic parameters such as sea state, direction, and velocity of the swell and the wind. It can also be seen as preprocessing that may be useful for target detection.; Cet article a pour but de fournir un algorithme efficace pour la segmentation du fouillis de mer dans les images radar cohérentes. En utilisant des modèles de mouvement d'interface à phase multiple, nous montrons que la segmentation dans la carte distance-temps-intensité donne des résultats beaucoup plus pertinents que la segmentation dans la carte distance-Doppler et évite l'utilisation d'hypothèses sur la réponse fréquentielle du fouillis. Plusieurs modèles mathématiques sont évalués sur des données réelles correspondant à des états de mer de Douglas d'environ 5. Le traitement de modèles de mouvement d'interface diffus impliquant plus que deux phases nous permet de prendre en compte le faible contraste, le bruit parasite et le manque de précision des contours qui sont typiques des images de la carte distance-temps-intensité. La segmentation du fouillis de mer nous permet, par exemple, d'obtenir des paramètres océanographiques tels que l'état de mer, la direction et la vitesse de la houle et du vent. Elle peut également être considérée comme un prétraitement qui peut être utile pour la détection de cibles.

Background subtraction is a fundamental pre-processing task in computer vision. This task becomes challenging in real scenarios due to variations in the background for both static and moving camera sequences. Several deep learning methods for background subtraction have been proposed in the literature with competitive performances. However, these models show performance degradation when tested on unseen videos; and they require huge amount of data to avoid overfitting. Recently, graph-based algorithms have been successful approaching unsupervised and semi-supervised learning problems. Furthermore, the theory of graph signal processing and semi-supervised learning have been combined leading to new insights in the field of machine learning. In this paper, concepts of recovery of graph signals are introduced in the problem of background subtraction. We propose a new algorithm called Graph BackGround Subtraction (GraphBGS), which is composed of: instance segmentation, background initialization, graph construction, graph sampling, and a semi-supervised algorithm inspired from the theory of recovery of graph signals. Our algorithm has the advantage of requiring less labeled data than deep learning methods while having competitive results on both: static and moving camera videos. GraphBGS outperforms unsupervised and supervised methods in several challenging conditions on the publicly available Change Detection (CDNet2014), and UCSD background subtraction databases.

Sports video analysis is a prevalent research topic due to the variety of application areas, ranging from multimedia intelligent devices with user-tailored digests up to analysis of athletes' performance. The Sports Video task is part of the MediaEval 2021 benchmark. This task tackles fine-grained action detection and classification from videos. The focus is on recordings of table tennis games. Running since 2019, the task has offered a classification challenge from untrimmed video recorded in natural conditions with known temporal boundaries for each stroke. This year, the dataset is extended and offers, in addition, a detection challenge from untrimmed videos without annotations. This work aims at creating tools for sports coaches and players in order to analyze sports performance. Movement analysis and player profiling may be built upon such technology to enrich the training experience of athletes and improve their performance., Comment: MediaEval 2021, Dec 2021, Online, Germany

Mathematical tools, machine learning, and signal processing tools have achieved enormous success in computer vision. In this chapter, we present state-of-the-art progress that has occurred in moving objects detection, classification, and recognition in video sequences taken by fixed cameras. More specifically, we focus on the latest breakthroughs made by deep learning (DL) that can be used for smart cities. The corresponding computer vision pipeline allows us to detect humans and vehicles for smart homes and cities. First, we survey developments in the field of moving objects detection using background subtraction techniques. Second, approaches that have been proposed to classify the extracted moving objects are also reviewed and classified into supervised, semi-supervised, and unsupervised learning methods. More specifically, they cover convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory (LSTM), autoencoder (AE), deep belief networks (DBN), and generative adversarial networks (GANs). Third, this chapter reviews approaches in the field of face recognition to identify people previously extracted using background subtraction and classified using deep neural networks (DNNs). Thus, this chapter aims to provide a comprehensive review of the segmentation, classification, and recognition of moving objects (humans, vehicles, etc.), covering conventional and recent advances.

Deep Convolutional Neural Networks (DCNNs) are widely used to extract high-dimensional features in various image recognition tasks [1] and have shown significant performance in face recognition. However, accurate real-time face recognition remains a challenge, mainly due to the high computation cost associated with the use of DCNNs and the need to balance precision requirements with time and resource restrictions. Besides, the supervised training process of DCNNs requires a large number of labeled samples. Aiming at solving the problem of data insufficiency, this study proposes a Deep Convolutional Generative Adversarial Net (DCGAN) based solution to increase the face dataset by generating synthetic images. Our proposed face recognition approach is based on FaceNet model. First, we perform face detection using MTCNN. After, a 128-D face embedding is extracted to quantify each face and a Support Vector Machine (SVM) is applied on top of the embeddings to recognize faces. In the experiment part, both LFW database and Chokepoint video database showed that our proposed approach with DCGANs data augmentation has improved the face recognition performance.

Dynamic Background Subtraction (BS) is a fundamental problem in many vision-based applications. BS in real complex environments has several challenging conditions like illumination variations, shadows, camera jitters, and bad weather. In this study, we aim to address the challenges of BS in complex scenes by exploiting conditional least squares adversarial networks. During training, a scene-specific conditional least squares adversarial network with two additional regularizations including L 1 -Loss and Perceptual-Loss is employed to learn the dynamic background variations. The given input to the model is video frames conditioned on corresponding ground truth to learn the dynamic changes in complex scenes. Afterwards, testing is performed on unseen test video frames so that the generator would conduct dynamic background subtraction. The proposed method consisting of three loss-terms including least squares adversarial loss, L 1 -Loss and Perceptual-Loss is evaluated on two benchmark datasets CDnet2014 and BMC. The results of our proposed method show improved performance on both datasets compared with 10 existing state-of-the-art methods.

In this article, a novel pixel based object detection framework is proposed that leverages dual type pixel-level information to construct the background model. The first type of information is initially used intensity histograms over a training set of a few initial video frames. Finally, it is formed by gathering all the minimum and maximum values of contiguous non-zero frequencies of the temporal intensity histogram. The second type of information constitutes a set having only the discrete pixel values. Subsequently, a pixel-level periodic updating scheme is used to make the model robust and flexible enough to recognize and detect foregrounds in various critical background environments. This dual format model produces effective results over many state-of-the-art methods in a large variety of challenging real-life video sequences.

This paper is devoted to an optimal control problem of the underground water contaminated by agricultural pollution, the spatiotemporal objective taking into account the trade-off between the fertilizer used by the farmer to increase profits and the cleaning costs which are necessary to treat the water before it is distributed to users. The constraint is a hydrogeological model for the spread of the pollution in the aquifer which consists in a system of a parabolic partial differential equation and an elliptic equation. Hydrogeological and economic modelling are thus combined in the problem. We propose a finite volume scheme based on a two-point flux approximation with upwind mobilities of an optimal control. Numerical simulations are provided to illustrate the 2D and 3D optimal solutions.

In this study, the authors present a new approach to segment and classify moving objects in video sequences by combining an unsupervised anomaly discovery framework called DeepSphere and generative adversarial networks. The proposed deep detector classifier employs and validates DeepSphere, which aims mainly to identify the anomalous cases in the spatial and temporal context in order to perform foreground objects segmentation. For post-processing, some morphological operations are considered to better segment and extract the desired objects. Finally, they take advantage of the power of generative models, which recognise the problem of semi-supervised learning as a specific missing data imputation task in order to classify the segmented objects. They evaluate the method with multiple datasets and the results confirm the effectiveness of the proposed approach, which achieves superior performance over the state-of-the-art methods having the capabilities of segmenting and classifying moving objects from videos surveillance.

In this paper, we present a new class of efficient models for water flow in shallow unconfined aquifers, providing an alternative to the classical but less tractable 3d-Richards model. Its derivation is guided by two objectives: to obtain a model that has low computational cost and yields relevant results on every time scale. Thus, we keep track of two types of flow that occur in such a context and are dominant when the ratio of thickness to longitudinal length is small: the first is dominant on a small time scale and is described by a vertical 1d-Richards problem; the second corresponds to a large time scale, when the evolution of the hydraulic head becomes independent of the vertical variable. These two types of flow are appropriately modeled by a one-dimensional and two-dimensional system of PDE boundary value problems, respectively. They are coupled at an artificial level below which the Dupuit hypothesis holds true (i.e., the vertical flow is instantaneous) so that the global model is mass conservative. Tuning the artificial level, which can even depend on an unknown of the problem, we obtain the new class of models. Using asymptotic expansions, we prove that the 3d-Richards model and each model in the class behave identically on every considered time scale (short, intermediate, and large) in thin aquifers. The results are illustrated by numerical simulations, and it is demonstrated that they fit well with those obtained by the original 3d-Richards model even in non-thin aquifers.

International audience; Human action recognition in video is one of the key problems in visual data interpretation. Despite intensive research, the recognition of actions with low inter-class variability remains a challenge. This paper presents a new Siamese Spatio-Temporal Convolutional Neural Network (SSTCNN) for this purpose. When applied to table tennis, it is possible to detect and recognize 20 table tennis strokes. The model has been trained on a specific dataset, so called TTStroke-21, recorded in natural conditions at the Faculty of Sports of the University of Bordeaux. Our model takes as inputs a RGB image sequence and its computed residual Optical Flow. The proposed siamese network architecture comprises 3 spatio-temporal convolutional layers, followed by a fully connected layer where data are fused. Our method reaches an accuracy of 91.4% against 43.1% for our baseline.

International audience

International audience

We propose a mathematical description of human color perception that relies on a hyper-bolic structure of the space P of perceived colors. We show that hyperbolicity allows us to reconcile both trichromaticity, from a Riemannian point of view, and color opponency, from a quantum viewpoint. In particular, we will underline how the opponent behavior can be represented by a rebit, a real analog of a qubit, whose state space is endowed with the Hilbert metric.

Computer vision applications based on videos often require the detection of moving objects in their first step. Background subtraction is then applied in order to separate the background and the foreground. In literature, background subtraction is surely among the most investigated field in computer vision providing a big amount of publications. Most of them concern the application of mathematical and machine learning models to be more robust to the challenges met in videos. However, the ultimate goal is that the background subtraction methods developed in research could be employed in real applications like traffic surveillance. But looking at the literature, we can remark that there is often a gap between the current methods used in real applications and the current methods in fundamental research. In addition, the videos evaluated in large-scale datasets are not exhaustive in the way that they only covered a part of the complete spectrum of the challenges met in real applications. In this context, we attempt to provide the most exhaustive survey as possible on real applications that used background subtraction in order to identify the real challenges met in practice, the current used background models and to provide future directions. Thus, challenges are investigated in terms of camera (i.e CCD cameras, omnidirectional cameras, …), foreground objects and environments. In addition, we identify the background models that are effectively used in these applications in order to find potential usable recent background models in terms of robustness, time and memory requirements.

During about 30 years, a lot of research teams have worked on the big challenge of detection of moving objects in various challenging environments. First applications concern static cameras but with the rise of the mobile sensors studies on moving cameras have emerged over time. In this survey, we propose to identify and categorize the different existing methods found in the literature. For this purpose, we propose to classify these methods according to the choose of the scene representation: one plane or several parts. Inside these two categories, the methods are grouped according to eight different approaches: panoramic background subtraction, dual cameras, motion compensation, subspace segmentation, motion segmentation, plane+parallax, multi planes and split image in blocks. A reminder of methods for static cameras is provided as well as the challenges with both static and moving cameras. Publicly available datasets and evaluation metrics are also surveyed in this paper., Comment: Submitted to Computer Science Review

Dynamic Background Modeling (DBM) is a crucial task in many computer vision based applications such as human activity analysis, traffic monitoring, surveillance, and security. DBM is extremely challenging in scenarios like illumination changes, camouflage, intermittent object motion or shadows. In this study, we proposed an end-to-end framework based on Generative Adversarial Network, which can generate dynamic background information for the task of DBM in an unsupervised manner. Our proposed model can handle the problem of DBM in the presence of the challenges mentioned above by generating data similar to the desired information. The primary aim of our proposed model during training is to learn all the dynamic changes in a scene-specific background information. While, during testing, inverse mapping of data to latent space representation in our model generates dynamic backgrounds similar to test data. The comparative analysis of our proposed model upon experimental evaluations on SBM.net and SBI benchmark datasets has outperformed eight existing methods for DBM in many challenging scenarios.

Action recognition in videos is one of the key problems in visual data interpretation. Despite intensive research, differencing and recognizing similar actions remains a challenge. This thesis deals with fine-grained classification of sport gestures from videos, with an application to table tennis.In this manuscript, we propose a method based on deep learning for automatically segmenting and classifying table tennis strokes in videos. Our aim is to design a smart system for students and teachers for analyzing their performances. By profiling the players, a teacher can therefore tailor the training sessions more efficiently in order to improve their skills. Players can also have an instant feedback on their performances.For developing such a system with fine-grained classification, a very specific dataset is needed to supervise the learning process. To that aim, we built the “TTStroke-21” dataset, which is composed of 20 stroke classes plus a rejection class. The TTStroke-21 dataset comprises video clips of recorded table tennis exercises performed by students at the sport faculty of the University of Bordeaux - STAPS. These recorded sessions were annotated by professional players or teachers using a crowdsourced annotation platform. The annotations consist in a description of the handedness of the player and information for each stroke performed (starting and ending frames, class of the stroke). Fine-grained action recognition has some notable differences with coarse-grained action recognition. In general, datasets used for coarse-grained action recognition, the background context often provides discriminative information that methods can use to classify the action, rather than focusing on the action itself. In fine-grained classification, where the inter-class similarity is high, discriminative visual features are harder to extract and the motion plays a key role for characterizing an action.In this thesis, we introduce a Twin Spatio-Temporal Convolutional Neural Network. This deep learning network takes as inputs an RGB image sequence and its computed Optical Flow. The RGB image sequence allows our model to capture appearance features while the optical flow captures motion features. Those two streams are processed in parallel using 3D convolutions, and fused at the last stage of the network. Spatio-temporal features extracted in the network allow efficient classification of video clips from TTStroke-21. Our method gets an average classification performance of 87.3% with a best run of 93.2% accuracy on the test set. When applied on joint detection and classification task, the proposed method reaches an accuracy of 82.6%.A systematic study of the influence of each stream and fusion types on classification accuracy has been performed, giving clues on how to obtain the best performances. A comparison of different optical flow methods and the role of their normalization on the classification score is also done. The extracted features are also analyzed by back-tracing strong features from the last convolutional layer to understand the decision path of the trained model. Finally, we introduce an attention mechanism to help the model focusing on particular characteristic features and also to speed up the training process. For comparison purposes, we provide performances of other methods on TTStroke-21 and test our model on other datasets. We notice that models performing well on coarse-grained action datasets do not always perform well on our fine-grained action dataset.The research presented in this manuscript was validated with publications in one international journal, five international conference papers, two international workshop papers and a reconductible task in MediaEval workshop in which participants can apply their action recognition methods to TTStroke-21. Two additional international workshop papers are in process along with one book chapter., La reconnaissance des actions à partir de vidéos est l'un des principaux problèmes de vision par ordinateur. Malgré des recherches intensives, la différenciation et la reconnaissance d'actions similaires restent un défi. Cette thèse porte sur la classification des gestes sportifs à partir de vidéos, avec comme cadre applicatif le tennis de table. Nous proposons une méthode d’apprentissage profond pour segmenter et classifier automatiquement les différents coup de Tennis de Table. Notre objectif est de concevoir un système intelligent permettant d'analyser les performances des élèves pongistes, et de donner la possibilité à l’entraîneur d'adapter ses séances d'entraînement pour améliorer leurs performances.Dans ce but, nous avons élaboré la base de données “TTStroke-21”, constituée de clips vidéo d'exercices de tennis de table, enregistrés par les étudiants de la faculté de sport de l'Université de Bordeaux – STAPS. Cette base de données a ensuite été annotée par des professionnels du domaine à l'aide d'une plateforme crowdsourcing. Les annotations consistent en une description des coups effectués (début, fin et type de coup). Au total, 20 différents coups de tennis de table sont considérés plus une classe de rejet. La reconnaissance des actions similaires présente des différences avec la reconnaissance d’actions classique. En effet, dans les bases de données classiques, le contexte de l’arrière plan fournit souvent des informations discriminantes que les méthodes peuvent utiliser pour classer l'action plutôt que de se concentrer sur l'action elle-même. Dans notre cas, la similarité entre classes est élevée, les caractéristiques visuelles discriminantes sont donc plus difficiles à extraire et le mouvement joue un rôle clef dans la caractérisation de l’action.Dans cette thèse, nous introduisons un réseau de neurones spatio-temporel convolutif avec une architecture Jumelle. Ce réseau d'apprentissage profond prend comme entrées une séquence d'images RVB et son flot optique estimé. Les données RVB permettent à notre modèle de capturer les caractéristiques d'apparence tandis que le flot optique capture les caractéristiques de mouvement. Ces deux flux sont traités en parallèle à l'aide de convolutions 3D, et sont fusionnés à la dernière étape du réseau. Les caractéristiques spatio-temporelles extraites dans le réseau permettent une classification efficace des clips vidéo de TTStroke-21. Notre méthode obtient une performance de classification de 93.2% sur l'ensemble des données tests. Appliquée à la tâche jointe de détection et de classification, notre méthode atteint une précision de 82.6%.Nous étudions les performances en fonction des types de données utilisés en entrée et la manière de les fusionner. Différents estimateurs de flot optique ainsi que leur normalisation sont testés afin d’améliorer la précision. Les caractéristiques de chaque branche de notre architecture sont également analysées afin de comprendre le chemin de décision de notre modèle. Enfin, nous introduisons un mécanisme d'attention pour aider le modèle à se concentrer sur des caractéristiques discriminantes et aussi pour accélérer le processus d’entraînement. Nous comparons notre modèle avec d'autres méthodes sur TTStroke-21 et le testons sur d'autres ensembles de données. Nous constatons que les modèles fonctionnant bien sur des bases de données d’actions classiques ne fonctionnent pas toujours aussi bien sur notre base de données d'actions similaires.Les travaux présentés dans cette thèse ont été validés par des publications dans une revue internationale, cinq papiers de conférences internationales, deux papiers d’un workshop international et une tâche reconductible dans le workshop MediaEval où les participants peuvent appliquer leurs méthodes de reconnaissance d'actions à notre base de données TTStroke-21. Deux autres papiers de workshop internationaux sont en cours de préparation, ainsi qu'un chapitre de livre.

Steganography is a science which helps to hide secret data inside multimedia supports like image, audio and video files to ensure secure communication between two parts of a channel. Steganalysis is the discipline which detects the presence of data hidden by a steganographic algorithm. There are two types of steganalysis: targeted steganalysis and universal steganalysis. In targeted steganalysis, the steganographic algorithm used to hide data is known. In the case of universal steganalysis, the detection of hidden data doesn’t depend on any specific algorithm used in the process of steganography. In this paper, we focus on universal steganalysis of images in a database with an eventual cover-source mismatch problem. It is shown that combining both unsupervised and supervised machine learning algorithms helps to improve the performance of classifiers in the case of universal steganalysis by reducing the cover-source mismatch problem. In the unsupervised step, the k-means algorithm is generally used to group similar images. When the number of features extracted from the image is very large it becomes difficult to compute the k-means algorithm properly. We propose, in that case, to use Deep Learning with Convolutional Neural Network (CNN) to group similar images at first and implement a Multilayer Perceptron (MLP) neural network to estimate the hidden message length in all the different groups of images. The first step of this approach prevents the cover-source mismatch problem. Reducing this issue boost the performance of classifiers in the second step which consists of estimating the hidden message length.

International audience; This paper presents a survey on the latest methods of moving object detection in video sequences captured by a moving camera. Although many researches and excellent works have reviewed the methods of object detection and background subtraction for a fixed camera, there is no survey which presents a complete review of the existing different methods in the case of moving camera. Most methods in this field can be classified into four categories: modelling based background subtraction, trajectory classification, low rank and sparse matrix decomposition, and object tracking. We discuss in details each category and present the main methods which proposed improvements in the general concept of the techniques. We also present challenges and main concerns in this field as well as performance metrics and some benchmark databases available to evaluate the performance of different moving object detection algorithms.

International audience; Iridescence is an original type of colouration that is relatively widespread in nature but has been either incompletely described or entirely neglected in prokaryotes. Recently, we reported a brilliant 'pointillistic' iridescence in agar-grown colony biofilms of Cellulophaga lytica and some other marine Flavobacteria that exhibit gliding motility. Bacterial iridescence is created by a unique self-organization of sub-communities of cells, but the mechanisms underlying such living photonic crystals are unknown. In this study, we used Petri dish assays to screen a large panel of potential activators or inhibitors of C. lytica's iridescence. Derivatives potentially interfering with quorum-sensing and other communication or biofilm formation processes were tested, as well as metabolic poisons or algal exoproducts. We identified an indole derivative, 5-hydroxyindole (5HI, 250 µM) which inhibited both gliding and iridescence at the colonial level. 5HI did not affect growth or cell respiration. At the microscopic level, phase-contrast imaging confirmed that 5HI inhibits the gliding motility of cells. Moreover, the lack of iridescence correlated with a perturbation of self-organization of the cell sub-communities in both the WT and a gliding-negative mutant. This effect was proved using recent advances in machine learning (deep neuronal networks). In addition to its effect on colony biofilms, 5HI was found to stimulate biofilm formation in microplates. Our data are compatible with possible roles of 5HI or marine analogues in the eco-biology of iridescent bacteria

In the last decades, researchers in the field of Background Subtraction (BS) have developed methods to handle the different type of challenges. However, at the present time, no traditional algorithm seems to be able to simultaneously address all the key BS challenges. This can mainly be attributed to the lack of systematic investigation concerning the role and the importance of features within background modeling and foreground detection. In this paper, we present a novel online one-class ensemble based on wagging to select suitable features to each region of a certain scene to distinguish the foreground objects from the background. In addition, we propose a mechanism to update the importance of each feature discarding insignificantly features over time. The experimental results on three challenging datasets (i.e MSVS, RGB-D object detection, CD.net 2014) show the pertinence of the proposed approach.

Inspired by the work of Resnikoff, which is described if full details in the first part of this two-part paper, we give a quantum description of the space P of perceived colors. We show that P is the effect space of a rebit, a real quantum qubit, whose state space S is isometric to the hyperbolic Klein disk K. This chromatic state space of perceived colors can be represented as a Bloch disk of real dimension 2 that coincides with the Hering disk given by the color opponency mechanism. Attributes of perceived colors, hue and saturation, are defined in terms of Von Neumann entropy.

Complete moving object detection plays a vital role in many applications of computer vision. For instance, depth estimation, scene understanding, object interaction, semantic segmentation, accident detection and avoidance in case of moving vehicles on a highway. However, it becomes challenging in the presence of dynamic backgrounds, camouflage, bootstrapping, varying illumination conditions, and noise. Over the past decade, robust subspace learning based methods addressed the moving objects detection problem with excellent performance. However, the moving objects detected by these methods are incomplete, unable to generate the occluded parts. Indeed, complete or occlusion-free moving object detection is still challenging for these methods. In the current work, we address this challenge by proposing a conditional Generative Adversarial Network (cGAN) conditioned on non-occluded moving object pixels during training. It therefore learns the subspace spanned by the moving objects covering all the dynamic variations and semantic information. While testing, our proposed Complete cGAN (CcGAN) is able to generate complete occlusion free moving objects in challenging conditions. The experimental evaluations of our proposed method are performed on SABS benchmark dataset and compared with 14 state-of-the-art methods, including both robust subspace and deep learning based methods. Our experiments demonstrate the superiority of our proposed model over both types of existing methods.

International audience; This essay inquires how mathematical beings could be inserted into the architecture of modes of existence proposed by Bruno Latour in the framework of his pluralist and renewed ontology of the modern world. An answer to this question is put forward with the aid of the work of Reviel Netz on the emergence of Greek mathematics, and of Charles Sanders Peirce on the diagrammatic dimension of mathematical practice, in the framework of an empiric conception of mathematics based on the notion of experience according to William James, and inspired by certain aspects of Per Martin-Löf's philosophy. It provides a way of describing the firm certainty with which proofs endow theorems, while invalidating the interpretation of this certainty as the mark of a direct access to an absolute and transcendental truth.; L'objet de cet essai est l'accueil des entités mathématiques dans l'architecture des modes d'existence proposée par Bruno Latour dans le cadre de son ontologie pluraliste du monde moderne. Les travaux de Reviel Netz sur l'émergence des mathématiques grecques et de Charles Sanders Peirce sur la dimension diagrammatique de l'activité mathématique sont employés pour proposer une réponse dans le cadre d'une conception empirique des mathématiques basée sur la notion d'expérience chère à William James et inspirée par certains aspects de la philosophie de Per Martin-Löf. Cette approche permet de penser la solide certitude dont la démonstration dote les résultats mathématiques tout en invalidant son interprétation comme la marque d'un accès direct à une vérité absolue et transcendante.

International audience

International audience; Detecting and classifying human actions in videos is oneof the current challenges in visual content analysis and mining. This paper presents a method for performing a fine-grained classification of sport actions using a Siamese Spatio-Temporal Convolutional Neural Network (SSTCNN) model.This model takes RGB images and Optical Flow field as in-put data. Our first contribution is the comparison of different Optical flow methods and a study of their influence on the classification score. We also present different normalization methods for the optical flow that drastically impact results,boosting performances from 44% to 74% of accuracy. Our second contribution is the detection and classification of actions in videos performed using a sliding temporal window.It leads to a satisfying score of 81.3% over the whole dataset TTStroke-21.

International audience; Human action recognition in videos is one of the key problems in visual data interpretation. Despite intensive research, the recognition of actions with low inter-class variability remains a challenge. To answer this problem, my thesis focus on fine-grained classification challenge using a Siamese Spatio-Temporal Convolutional Neural Network and apply it to a new dataset we have introduced TTStroke-21. Our model take as input data RGB images and Optical Flow and is able to reach an accuracy of 91.4% against 43.1% for our baseline on temporal segmented videos. Detection and classification in videos using a sliding temporal window leads to a score of 81.3% over the whole dataset.

In this thesis, we are interested in the problem of 3D segmentation of skin tumors in high frequency ultrasound images. We focus mainly on two questions: how best to estimate the volume of tumors (in accordance with references produced by dermatologists) and how to produce algorithms whose computation times are close to real time? First, we describe a new model, log-likelihood Cahn-Hilliard (LLCH), based on a variational formulation coupling a data attachment term computed from non-parametric estimates and a regularization term derived from a phase transition dynamic (Allen Cahn reaction diffusion equation). This model is tested with a first multigrid implementation using exact solutions calculated with a Lie splitting. Secondly, we are interested in the possibility of implementing the LLCH model using lattice Boltzmann methods (LBM). The underlying dynamic is not physical in nature, so this implementation is not direct and is subject to instability problems. We show that, due to the specificities of the data attachment term, the BGK schemes, with simple relaxation time, do not ensure sufficient stability. We then use MRT schemes, with multiple relaxation times, which allow us to gain stability by introducing additional parameters. The adjustment of the so-called quartic parameters makes it possible to obtain fourth-order exact schemes that are numerically stable. Tests performed on a clinical database with ground truth provided by dermatologists show that the results obtained with the two proposed implementations are much better than those obtained with level sets methods and that our model is a good alternative to overcome the problem of underestimation of tumor volume. The computation times, for 3D images of about 70 million voxels, are very short and well adapted for practical use in medical environments., Nous nous intéressons dans cette thèse au problème de la segmentation 3D de tumeurs de la peau dans des images ultrasons hautes fréquences. Nous nous concentrons essentiellement sur deux questions : comment estimer au mieux le volume des tumeurs (en accord avec les références produites par des dermatologues) et comment produire des algorithmes dont les temps de calcul se rapprochent du temps réel ? Dans un premier temps, nous décrivons un nouvel modèle, log-likelihood Cahn-Hilliard (LLCH), basé sur une formulation variationnelle couplant un terme d’attache aux données calculé à partir d’estimations non paramétriques et un terme de régularisation issu d’une dynamique de transitions de phase (équation de réaction diffusion d’Allen Cahn). Ce modèle est testé avec une première implémentation multigrille par solutions exactes calculées grâce à un splitting de Lie. Dans un second temps, nous nous intéressons à la possibilité d’implémenter le modèle LLCH par des méthodes Lattice Boltzmann (LBM). La dynamique sous-jacente n’étant pas de nature physique, cette implémentation n’est pas directe et est sujette à des problèmes d’instabilité. Nous montrons que, compte tenu des spécificités du terme d’attache aux données, les schémas BGK, à simple temps de relaxation, ne permettent pas d’assurer une stabilité suffisante. Nous avons alors recours à des schémas MRT, à temps de relaxation multiples, qui permettent par l’introduction de paramètres additionnels de gagner en stabilité. L’ajustement des paramètres dits quartiques permet d’obtenir des schémas exacts à l’ordre 4 et numériquement stables. Les tests réalisés sur une base de données cliniques avec une vérité terrain fournie par des dermatologues montrent que les résultats obtenus grâce aux deux implémentations proposées sont bien meilleurs que ceux obtenus par les méthodes level sets et que notre modèle est une bonne alternative pour pallier le problème de la sous-estimation du volume tumoral. Les temps de calcul, pour des images 3D d’environ 70 millions de voxels, sont très courts et tout-à-fait adaptés pour une utilisation pratique en milieu médical.

International audience; This project explores a new collaborative approach for education in medical imaging diagnostic. This approach is co-adaptative between interaction (human) and automation (computation).; Le projet concerne l'exploration de nouveaux modes d'apprentissage collaboratifs, suivant une approche co-adaptative : interactive (humaine) et automatique (calculée), appliquée à l'imagerie médicale endoscopique.

The aim of this paper is to present a new method for skin tumor segmentation in the 3D ultrasound images. We consider a variational formulation, the energy of which combines a diffuse interface phase field model (regularization term) and a log-likelihood computed using nonparametric estimates (data attachment term). We propose a multi-grid implementation with the exact solutions which has the advantage to avoid space discretization and numerical instabilities. The resulting algorithm is simple and easy to implement in multi-dimensions. Concerning applications, we focus on skin tumor segmentation. The clinical dataset used for the experiments is composed of 12 images with the ground truth given by a dermatologist. Comparisons with the reference methods show that the proposed method is more robust to the choice of the volume initialization. Moreover, thanks to the flexibility introduced by the diffuse interface, the sensitivity increases by 12% if the initialization is inside the lesion, and the Dice index increases by 59%, if the initialization covers the entire lesion. These results show that this new method is well designed to tackle the problem of underestimation of tumor volumes.

International audience; This paper presents a new construction of a homogeneous Dirichlet wavelet basis on the unit interval, linked by a diagonal differentiation-integration relation to a standard biorthogonal wavelet basis. This allows to compute the solution of Poisson equation by renormalizing the wavelet coefficients - as in Fourier domain but using locally supported basis functions with boundary conditions-, which yields a linear complexity $O(N)$ for this problem. Another application concerns the construction of free-slip divergence-free wavelet bases of the hypercube, in general dimension, with an associated decomposition algorithm as simple as in the periodic case.