Human pose estimation via multi-layer composite models

We introduce a hierarchical part-based approach for human pose estimation in static images. Our model is a multi-layer composite of tree-structured pictorial-structure models, each modeling human pose at a different scale and with a different graphical structure. At the highest level, the submodel acts as a person detector, while at the lowest level, the body is decomposed into a collection of many local parts. Edges between adjacent layers of the composite model encode cross-model constraints. This multi-layer composite model is able to relax the independence assumptions in tree-structured pictorial-structures models (which can create problems like double-counting image evidence), while still permitting efficient inference using dual-decomposition. We propose an optimization procedure for joint learning of the entire composite model. Our approach outperforms the state-of-the-art on four challenging datasets: Parse, UIUC Sport, Leeds Sport Pose and FLIC datasets.