Probabilistic Characterization of Pose Uncertainty under Self Occlusion

In this paper, we study the uncertainty in human body pose from 2D silhouette observations. Under perspective imaging, the loss of depth information and self-occlusion of the articulating limbs result in multiple 3D poses mapping on to similar 2D silhouettes. This introduces ambiguity in using the silhouette for inference, particularly in capturing human motion. To characterize this uncertainty, we present an algorithm which approximates the probability map from a silhouette to its possible 2D poses by sampling efficiently on a 2D pose space. This algorithm is driven using a articulated human body model as a prior and does not require any training, or prior assumption on the motion. When a temporal observation sequence is available (such as a video) temporal smoothness can be used to further prune the uncertainty space. Further, we show that the probability maps from different camera views can be geometrically fused to obtain the pose uncertainty distribution in 3D space. Finally, the entropy of the posterior distribution is used as a measure for quantizing the amount of information encoded by the silhouette. This has applications in perceptual grouping of silhouettes and in view selection for multi-camera networks. We also show that the pose uncertainty decreases with number of camera views as well as consecutive frame observations from video. We also verified the pose estimation (both single and multi-view) over standard motion capture datasets.