Silhouette representation and matching for 3D pose discrimination – A comparative study

Inferring 3D human poses from marker-free images is an important but challenging task. A large body of algorithms has been proposed to that end, among which the discriminative methods using silhouettes as visual inputs are an important category. For these methods, silhouette representation and matching is very important. An effective silhouette representation method computes discriminative and compact silhouette descriptors which are used for learning the silhouette-pose mapping, and a good silhouette matching algorithm enables effective comparison and search in the example database. However, there has not been an extensive study on the abundance of shape analysis techniques in the context of pose discrimination. In this paper, we give a systematic study on the performances of shape representation and matching algorithms for pose discrimination, and we explore the influences of different realistic factors encountered in practical systems, such as yaw angle, camera tilt, silhouette noise, and the selection of training examples. We conduct various quantitative evaluations using synthetic and real silhouettes based on HumanEva dataset. Our work provides new insights into pose inferring algorithms and the designing and building of practical systems.