Search

International audience; This paper deals with the binary classification of seals videos. To address this problem, we propose a novel system based on two phases : offline and online. The offline phase aims to generate a trained model. For this end, we use a transfer learning of convolutional neural network approach.Especially, we employed the VGG-16 deep network trained on the large ImageNet database and transfer their hyperparameters torecognize seals in the video. Finally, fully connected layer followed by a softmax method are used to decide if the seals exist ornot in a given frame. For the online phase, we extract different information from a given video based on the file video name and on the trained model in the offline phase. The evaluation of the proposed method in a seals videos database demonstrates promising results.

In this article, we present a new method of dehazing based on the Koschmieder model, which aims to restore an image that has been affected by haze. The difficulty is to improve the estimation of the transmission and the atmospheric light that generally suffer from the nonhomogeneity and the random variability of the environment. The keypoint is to enhance the dehazing of very bright regions of the image in order to improve the treatment of the sky that is often overestimated or underestimated compared to the rest of the scene. The approach proposed in this paper is based on two main contributions: 1. an L0 gradient optimization function weighted by a set of Gaussian filters and based on an iterative algorithm for optimization convergence. Unlike the existing methods using a single value of the atmospheric light for the whole image, our method uses a set of values neighboring an initial estimated value. The fusion is then applied based on Laplacian and Gaussian pyramids to combine all the relevant information from the set of images constructed from atmospheric lights and improves the contrast to recover the colors of the sky without any artifacts. Finally, the results are validated by three criteria: an autocorrelation score (ZNCC), a similarity measure (SSIM) and a visual criterion. The experiments carried out on two datasets show that our approach allows a better dehazing of the images with higher SSIM and ZNCC measurements but also with better visual quality.

In this paper, we propose an adaptation of marker detection algorithm for panoramic cameras such as catadioptric and fisheye sensors. Due to distortions and non-uniform resolution of such sensors, the methods that are commonly used in perspective images cannot be applied directly. This work is in contrast with the existing marker detection framework: Automatic reliable fiducial markers Under occlusion (ArUco) for a conventional camera. To keep the same performance for panoramic cameras, our method is based on a spherical representation of the image that allows the marker to be detected and to estimate its 3D pose. We evaluate our approach on a new shared dataset that consists of a 3D rig of markers taken with two different sensors: a catadioptric camera and a fisheye camera. The evaluation has been performed against ArUco algorithm without rectification and with one of the rectified approaches based on the fisheye model.

International audience; Most of dehazing algorithms are based on the well known Koschmieder image restoration model working on color images. Nowadays, many autonomous underwater vehicles use marker detection in order to estimate their 3D pose according to the marked target. In this paper, we propose an adaptation of the Koschmieder for grayscale images which is more suitable for marker detection. This is done by enhancing the final energy (the radiance) of the image lost during the processing. We show that by multiplying the final radiance image by the shifted transmission of the Koschmieder model, we can enhance the gradient and the contrast of the image. We have implemented our Koschmieder adaptation into two methods from the literature and proven its robustness on an underwater dataset containing ArUco markers. The obtained results outperform the existing methods in terms of marker detection rate without degrading the pose estimation.

International audience; In order to accompany the swimming coaches in evaluating high-level swimmers, we developed a prototype for instantaneous speed estimation. To achieve this, we proposed and validated, in a previous work, a swimmer tracking system based on data fusion. However, the initialization phase is done manually, and our aim, in this paper, is to automate this process. First, we propose a region of interest localization module that allows the detection of the first appearance of the swimmer in the lane as well as the restriction of the region of interest around him. This module is based on the method a contrario which consists of modeling the random noise corresponding to the water and detecting the structured movement relative to the swimmer motion. To do that, we calibrate the pool using DLT (Direct Linear Transform) technique, extract the concerned lane, apply the frame difference approach to detect the moving objects, and then decompose the lane into blocs and classify them into swimmer motion or noise. Second, in order to detect the swimmer’s head, we propose the Scaled Composite JTC which is based on the NL-JTC correlation technique. The input plane of this latter includes a target and a reference image. The first is the region of interest detected by the method a contrario. The second consists of a Scaled Composite Reference. The tests conducted on real video sequences of French swimming championships (Limoges 2015) showed very good results in terms of region of interest localization and swimmer’s head detection which allows a reliable initialization for the tracking system.

Despite its extensive use, the traditional 4f Vander Lugt Correlator optical setup can be further simplified. We propose a lightweight correlation scheme where the decision is taken in the Fourier plane. For this purpose, the Fourier plane is adapted and used as a decision plane. Then, the offline phase and the decision metric are re-examined in order to keep a reasonable recognition rate. The benefits of the proposed approach are numerous: (1) it overcomes the constraints related to the use of a second lens; (2) the optical correlation setup is simplified; (3) the multiplication with the correlation filter can be done digitally, which offers a higher adaptability according to the application. Moreover, the digital counterpart of the correlation scheme is lightened since with the proposed scheme we get rid of the inverse Fourier transform (IFT) calculation (i.e., decision directly in the Fourier domain without resorting to IFT). To assess the performance of the proposed approach, an insight into digital hardware resources saving is provided. The proposed method involves nearly 100 times fewer arithmetic operators. Moreover, from experimental results in the context of face verification-based correlation, we demonstrate that the proposed scheme provides comparable or better accuracy than the traditional method. One interesting feature of the proposed scheme is that it could greatly outperform the traditional scheme for face identification application in terms of sensitivity to face orientation. The proposed method is found to be digital/optical implementation-friendly, which facilitates its integration on a very broad range of scenarios.

International audience; In this paper, an all-automatic optimized JTC based swimmer tracking system is proposed and evaluated on real video database outcome from national and international swimming competitions (French National Championship, Limoges 2015, FINA World Championships, Barcelona 2013 and Kazan 2015). First, we proposed to calibrate the swimming pool using the DLT algorithm (Direct Linear Transformation). DLT calculates the homography matrix given a sufficient set of correspondence points between pixels and metric coordinates: i.e. DLT takes into account the dimensions of the swimming pool and the type of the swim. Once the swimming pool is calibrated, we extract the lane. Then we apply a motion detection approach to detect globally the swimmer in this lane. Next, we apply our optimized Scaled Composite JTC which consists of creating an adapted input plane that contains the predicted region and the head reference image. This latter is generated using a composite filter of fin images chosen from the database. The dimension of this reference will be scaled according to the ratio between the head's dimension and the width of the swimming lane. Finally, applying the proposed approach improves the performances of our previous tracking method by adding a detection module in order to achieve an all-automatic swimmer tracking system.

International audience; To design a robust swimmer tracking system, we took into account two well-known tracking techniques: the nonlinear joint transform correlation (NL-JTC) and the color histogram. The two techniques perform comparably well, yet they both have substantial limitations. Interestingly, they also seem to show some complementarity. The correlation technique yields accurate detection but is sensitive to rotation, scale and contour deformation, whereas the color histogram technique is robust for rotation and contour deformation but shows low accuracy and is highly sensitive to luminosity and confusing background colors. These observations suggested the possibility of a dynamic fusion of the correlation plane and the color scores map. Before this fusion, two steps are required. First is the extraction of a sub-plane of correlation that describes the similarity between the reference and target images. This sub-plane has the same size as the color scores map but they have different interval values. Thus, the second step is required which is the normalization of the planes in the same interval so they can be fused. In order to determine the benefits of this fusion technique, first, we tested it on a synthetic image containing different forms with different colors. We thus were able to optimize the correlation plane and color histogram techniques before applying our fusion technique to real videos of swimmers in international competitions. Last, a comparative study of the dynamic fusion technique and the two classical techniques was carried out to demonstrate the efficacy of the proposed technique. The criteria of comparison were the tracking percentage, the peak to correlation energy (PCE), which evaluated the sharpness of the peak (accuracy), and the local standard deviation (Local-STD), which assessed the noise in the planes (robustness).

This paper describes a Web-based system that is aimed to help both academic staff and students in the task of automatic assessment of programming assignments. The server implements LAMP (Linux, Apache, MySQL and PHP) architecture powered by the Laravel Framework and designed by Twitter bootstrap. The system focuses on adaptability, scalability and simplicity. It also enables teachers with few skills in the field of software testing to use such a tool. This system implements various types of technologies: databases, Web technologies, object oriented programming, responsive Web design, design patterns, and others.

International audience; The face recognition tasks can be divided into two categories: verification (i.e. compare two images in order to know if they represent the same person) and identification (i.e. find the identity of a person into the data-base). Several powerful face recognition methods exist, in literature, for controlled environments: constrained illumination, frontal pose, neutral expression... However, there are few reliable methods for the uncontrolled case. Optical correlation has shown its interest through relevant architectures for controlled and uncontrolled environments. Based on this architecture, we propose a novel method for verification and identification tasks under illumination variation conditions. More specifically, we optimize the performances of a correlation method against illumination changes by using and adapting the Local Binary Patterns (LBP) description. This later is widely used in the literature to describe the texture of an image using 8 bits words. For both, target image and reference image, we begin by using a specific-Gaussian function as first step of LBP-VLC correlator. This function filters the considered image with a band-pass filter in order to extract the edges. Then we applied the adapted LBP-VLC method. To validate our new approach, we used a simple POF filter (others correlation filters can be used). The simulations are done using the YaleB and YaleB Extended databases that contain respectively 10 and 38 identities with 64 illuminations. The results obtained reach more than 94% and 92% for the verification and 93% and 90% for the identification case. These results show the good performances of our approach of LBP-correlation methods against illumination changes.

Poster session: WS21 - Workshop on Information Fusion in Computer Vision for Concept Recognition; International audience; The REPERE challenge is a project aiming at the evaluation of systems for supervised and unsupervised multimodal recognition of people in TV broadcast. In this paper, we describe, evaluate and discuss QCompere consortium submissions to the 2012 REPERE evaluation campaign dry-run. Speaker identification (and face recognition) can be greatly improved when combined with name detection through video optical character recognition. Moreover, we show that unsupervised multimodal person recognition systems can achieve performance nearly as good as supervised monomodal ones (with several hundreds of identity models).

3D object recognition and retrieval recently gained a big interest because of the limitation of the "2D-to-2D" approaches. The latter suffer from several drawbacks such as the lack of information (due for instance to occlusion), pose sensitivity, illumination changes, etc. Our main motivation is to gather both discrimination and easy interaction by allowing simple (but multiple) 2D specifications of queries and their retrieval into 3D gallery sets. We introduce a novel "2D sketch-to-3D model" retrieval framework with the following contributions: (i) first a novel generative approach for aligning and normalizing the pose of 3D gallery objects and extracting their 2D canonical views is introduced. (ii) Afterwards, robust and compact contour signatures are extracted using the set of 2D canonical views. We also introduce a pruning approach to speedup the whole search process in a coarseto- fine way. (iii) Finally, object ranking is performed using our variant of elastic dynamic programming which considers only a subset of possible matches thereby providing a considerable gain in performance for the same amount of errors. Our experiments are reported/compared through the Princeton Shape Benchmark; clearly showing the good performance of our framework w.r.t. the other approaches. An iPhone demo of this method is available and allows us to achieve "2D sketch to 3D object" querying and interaction.

We describe in this paper a method for 3D shape indexing and retrieval that we apply on three data collections of the SHREC - SHape Retrieval Contest 2008: Stability on watertight, CAD and Generic 3D models. The method is based on a set of 2D multi-views after a pose and scale normalization of the models using Continuous PCA and the enclosing sphere. In all views we extract the models silhouettes and compare them pairwise. To compute the similitude measure we consider the external contour of the silhouettes, we extract their convexities and concavities at different scale levels and we build a multiscale representation. The pairs of contours are then compared by elastic matching achieved by using dynamic programming.

International audience; We describe in this paper two methods for 3D shape indexing and retrieval that we apply on two data collections of the SHREC - SHape Retrieval Contest 2007: Watertight models and 3D CAD models. Both methods are based on a set of 2D multi-views after a pose and scale normalization of the models using PCA and the enclosing sphere. In all views we extract the models silhouettes and compare them pairwise. In the first method the similitude measure is obtained by integrating on the pairs of views the difference between the areas of the silhouettes union and the silhouettes intersection. In the second method we consider the external contour of the silhouettes, extract their convexities and concavities at different scale levels and build a multiscale representation. The pairs of contours are then compared by elastic matching achieved by using dynamic programming. Comparisons of the two methods are shown with their respective strengths and weaknesses.

In this thesis, we focus on issues related to 3D objects indexing and content based retrieval. Particularly, we focused on multi-view indexing methods. These approaches, characterized the shape by using 2D projections of the 3D object. First, we introduce a new approach to normalize and align the 3D objects necessary for our indexing process. The definition of position and scale is based on the minimal bounding sphere that offers interesting properties for our multi-view characterization of the shape. To find an optimal alignment for the objects, we propose an estimator based on results obtained in cognitive psychology to compare two different poses. In a second part, we define three new shape descriptors based on 2D projections. The first describes the shape of a silhouette with a set of pixels. This allows using set operators to compare signatures. In a second step, we use convexities and concavities information to describe the contour of the projections of our 3D objects. With these measures we have defined two related descriptors based on histograms and DCT compression. Finally, we propose a last descriptor where 2D projections are associated with orientation surface information. This "normal map" descriptor is compressed using Fourier coefficients. Finally, our retrieval process can query the database using 3D objects, pictures or sketches. To ensure effective search in time and relevance of results, we propose two optimizations. The first is based on results fusion using different aggregation operators while the latter quickly eliminates distant objects to the query through an early pruning.; Dans cette thèse, nous nous intéressons aux problématiques liées à l'indexation et à la recherche d'objets 3D par le contenu. Dans un cadre plus particulier, nous nous sommes penchés sur des méthodes d'indexation multi-vues, caractérisant la forme à l'aide de projections 2D de l'objet en trois-dimensions. Dans un premier temps, nous introduisons une nouvelle approche de normalisation et d'alignement des objets 3D nécessaire à notre processus d'indexation. La définition de la position et de l'échelle s'appuie sur la sphère minimale englobante qui offre des propriétés intéressantes pour notre caractérisation multi-vues de la forme. En ce qui concerne la recherche de l'alignement optimal, nous proposons un estimateur permettant de comparer deux poses différentes en s'appuyant sur des résultats obtenus en psychologie cognitive. Dans une seconde partie, nous définissons trois nouveaux descripteurs de forme basés sur des projections 2D. Le premier décrit la forme d'une silhouette à l'aide d'un ensemble de pixels. Le second, utilise des informations de convexités et de concavités pour décrire le contour des projections de nos objets 3D. Enfin, le dernier utilise des informations d'orientation de la surface 3D. Finalement, le processus de recherche que nous introduisons permet d'interroger la base de données à l'aide d'objets 3D, de photos ou de dessins au trait. Afin de garantir une recherche efficace en temps et en pertinence des résultats, nous proposons deux optimisations. La première s'appuie sur la fusion de résultats tandis que la seconde élimine rapidement les objets éloignés de la requête au moyen d'un élagage précoce.

International audience; This paper proposes a novel face representation based on Local Quantized Patterns (LQP). LQP is a generalization of local pattern features that makes use of vector quantization and lookup table to let local pattern features have many more pixels and/or quantization levels without sacrificing simplicity and computational efficiency. Our new LQP face representation not only outperforms any other representation on challenging face datasets but performs equally well in the intensity space and orientation space (obtained by applying gradient or Gabor Filters) and hence is intrinsically robust to illumination variations. Extensive experiments on several challenging face recognition datasets (such as FERET and LFW) show that this representation gives state-of-the-art performance (improving the earlier state-of-the-art by around 3%) without requiring neither a metric learning stage nor a costly labelled training dataset, having the comparison of two faces being made by simply computing the Cosine similarity between their LQP representations in a projected space.