Multi-view object categorization and pose estimation

Abstract

Object and scene categorization has been a central topic of computer vision research in recent years. The problem is a highly challenging one. A single object may show tremendous variability in appearance and structure under various photometric and geometric conditions. In addition, members of the same class may differ from each other due to various degrees of intra-class variability. Recently, researchers have proposed new models towards the goal of: i) finding a suitable representation that can efficiently capture the intrinsic three-dimensional and multi-view nature of object categories; ii) taking advantage of this representation to help the recognition and categorization task. In this Chapter we will review recent approaches aimed at tackling this challenging problem and focus on the work by Savarese & Fei-Fei [54, 55]. In [54, 55] multi-view object models are obtained by linking together diagnostic parts of the objects from different viewing point. Instead of recovering a full 3D geometry, parts are connected through their mutual homographic transformation. The resulting model is a compact summarization of both the appearance and geometry information of the object class. We show that such a model can be learnt via minimal supervision compared to competitive techniques. The model can be used to detect objects under arbitrary and/or unseen poses by means of a two-step algorithm. This algorithm, inspired by works in single object view synthesis (e.g., Seitz & Dyer [57]), has the ability to synthesize object appearance and shape properties at recognition time, and in turn estimate the object pose that best matches the observations.We conclude this Chapter by presenting experiments on detection, recognition and pose estimation results with respect to two datasets in [54,55] as well as to PASCAL Visual Object Classes (VOC) dataset [15]. Experiments indicate that representation and algorithms presented in [54,55] can be successfully employed in a number of generic object recognition tasks. © 2010 Springer-Verlag Berlin Heidelberg.