Segmentation de catégories d’objets par combinaison d’un modèle d’apparence et d’un champs de Markov

Object models based on bag-of-words representations achieve state-of-the-art performance for image classification and object localization. However, as they consider objects as lose collection of information they fail to accurately locate object boundaries and thus produce inaccurate object segmentation. On the other hand, Markov Random Field based models used for image segmentation foFIG. 1 – 4 exemples de segmentations obtenues par notre méthode. Les instances de catégories d’objets sont automatiquement localisées dans l’image, produisant des masques de segmentations qui peuvent être utilisés pour extraire automatiquement les objets. cus on object boundaries but can hardly use global object constraints, which is required when dealing with object categories whose appearance may vary significantly. The key contribution of this paper is to combine elegantly the advantages of these two approaches. First, a blob-based mechanism allows to detect objects using visual words occurrences, and produces rough image segmentation. Second, a MRF component produces clean cuts, guided by local image cues (color, texture and edge cues) and by longdistance dependency given by the blob model, which enforces label consistency. Gibbs sampling is used to infer the model. Experiments on Pascal VOC 2006 and Graz02 datasets show impressive results in the difficult context of object categories segmentation in presence of cluttered backgrounds and large view point changes.