2D and 3D Mondels for Object and Scene Understanding

Abstract

In this thesis, we propose novel 2D and 3D models for object and scene understanding from images. This is an extremely challenging problem in computer vision. Objects change their appearance because of intra-class variability, view point transformations and their inherent deformable nature. Understanding scenes is also challenging. Scenes may comprise large number of objects whose relationships are class specific and depend on the view point and 3D scene geometry. First, we propose object models that are capable of detecting generic rigid objects and simultaneously extracting their viewpoints and 3D shape. The ability of these models to jointly capture appearance and shape in a truly 3D sense makes them robust to intra-class variability, view point changes and occlusions. Second, we have focused on designing models that can effectively capture the appearance and shape variability of deformable objects such as humans or animals. Most importantly, we propose algorithmic solutions that are capable of “taming” the intrinsic complexity of the pose estimation problem while guaranteeing the optimality of the solution. Finally, we have proposed models that can effectively capture the interplay among objects and scene elements so as to simultaneously recognize the scene, detect objects and segment regions accurately and efficiently.