Contextual Rescoring for Human Pose Estimation

Given an image of a person, the problem of human pose estimation can be briefly described as localizing the position and orientation of the body limbs. The complexity of the problem comes from issues like background clutter, changes in viewpoint, changes in appearance, self-occlusions of body parts, etc. Pictorial structures framework has been widely applied in human pose estimationn during the past few years [1]. Yang and Ramanan [7] proposed a simple yet efficient model that outperformed previous state of the art approaches. However, in addition to the difficulties of modelling small image patches for the body joints (see Fig. 1), the performance of their method is also compromised by the use of a tree-structured model. Although trees permit efficient and exact inference on graphical models, the restricted edge structure is insufficient for capturing all the important relations between parts. As a consequence, tree-structured pictorial structures suffer from the so-called “double-counting” phenomena.