Abstract

Tree-structured models have been widely used for human pose estimation, in either 2D or 3D. While such models allow efficient learn- ing and inference, they fail to capture additional dependencies between body parts, other than kinematic constraints between connected parts. In this paper, we consider the use of multiple tree models, rather than a single tree model for human pose estimation. Our model can alleviate the limitations of a single tree-structured model by combining informa- tion provided across different tree models. The parameters of each indi- vidual tree model are trained via standard learning algorithms in a single tree-structured model. Different tree models can be combined in a discrim- inative fashion by a boosting procedure. We present experimental results showing the improvement of our approaches on two different datasets. On the first dataset, we use our multiple tree framework for occlusion rea- soning. On the second dataset, we combine multiple deformable trees for capturing spatial constraints between non-connected body parts.