Segmentation multi-vues d'objet

Keywords :

• analyse de scène
• suivi

There has been a growing interest for multi-camera systems and many interesting works have tried to tackle computer vision problems in this particular configuration. The general objective is to propose new multi-view oriented methods instead of applying limited monocular approaches independently for each viewpoint. The work in this thesis is an attempt to have a better understanding of the multi-view object segmentation problem and to propose an alternative approach making maximum use of the available information from different viewpoints. Multiple view segmentation consists in segmenting objects simultaneously in several views. Classic monocular segmentation approaches reason on a single image and do not benefit from the presence of several viewpoints. A key issue in that respect is to ensure propagation of segmentation information between views while minimizing complexity and computational cost. In this work, we first investigate the idea that examining measurements at the projections of a sparse set of 3D points is sufficient to achieve this goal. The proposed algorithm softly assigns each of these 3D samples to the scene background if it projects on the background region in at least one view, or to the foreground if it projects on foreground region in all views. A complete probabilistic framework is proposed to estimate foreground/background color models and the method is tested on various datasets from state of the art. Two different extensions of the sparse 3D sampling segmentation framework are proposed in two scenarios. In the first, we show the flexibility of the sparse sampling framework, by using variational inference to integrate Gaussian mixture models as appearance models. In the second scenario, we propose a study of how to incorporate depth measurements in multi-view segmentation. We present a quantitative evaluation, showing that typical color-based segmentation robustness issues due to color-space ambiguity between foreground and background, can be at least partially mitigated by using depth, and that multi-view color depth segmentation also improves over monocular color depth segmentation strategies. The various tests also showed the limitations of the proposed 3D sparse sampling approach which was the motivation to propose a new method based on a richer description of image regions using superpixels. This model, that expresses more subtle relationships of the problem trough a graph construction linking superpixels and 3D samples, is one of the contributions of this work. In this new framework, time related information is also integrated. With static views, results compete with state of the art methods but they are achieved with significantly fewer viewpoints. Results on videos demonstrate the benefit of segmentation propagation through geometric and temporal cues. Finally, the last part of the thesis explores the possibilities of tracking in uncalibrated multi-view scenarios. A summary of existing methods in this field is presented, in both mono-camera and multi-camera scenarios. We investigate the potential of using self-similarity matrices to describe and compare motion in the context of multi-view tracking.

Directors:

• Mr Patrick Perez (Directeur de Recherche - INRIA )
• Mr Edmond Boyer (Directeur de Recherche - INRIA )
• Mr Sébastien Franco (Maître de Conférence - Grenoble INP )

Raporteurs:

• Mr Adrian Hilton (Professeur - University of Surrey )
• Mr Daniel Cremers (Professeur - Technische Universität München )

Examinators:

• Mr George Drettakis (Directeur de recherche - INRIA Sophia Antipolis )