Synthetic 3D Model-Based Object Class Detection and Pose Estimation. (Détection de Classes d'Objets et Estimation de leurs Poses à partir de Modèles 3D Synthétiques)

The present thesis describes 3D model-based approaches to object class detection and pose estimation on single 2D images. We introduce learning, detection and estimation steps adapted to the use of synthetically rendered training data with known 3D geometry. Most existing approaches recognize object classes for a particular viewpoint or combine classifiers for a few discrete views. By using CAD models and rendering techniques from the domain of computer graphics, we propose instead to build 3D representations of object classes which allow to handle viewpoint changes and intra-class variability. We outline an unsupervised filtering process of pose and class discriminant local features on purely synthetic training data, and we derive a part model which discriminatively learns the object class appearance from an annotated database of real images and builds a generative representation of its 3D geometry from a database of synthetic CAD models. During detection, we introduce a 3D voting scheme to reinforce geometric coherence by means of a robust pose estimation, and we propose an alternative probabilistic method which evaluates the likelihood of groups of 2D part detections with respect to a full 3D geometry. Both approaches yield approximate 3D bounding boxes in addition to 2D localizations; these initializations are subsequently improved and disambiguated by a registration scheme aligning arbitrary 3D models to 2D images. The work is evaluated on several standard benchmark datasets and achieves state-of-the-art performance for 2D detection in addition to providing 3D pose estimations from single images.