Discriminative 3D human pose estimation from monocular images via topological preserving hierarchical affinity clustering

This paper presents a hierarchical approach to address the problem of 3D human body pose estimation from a single images. In order to deal with multimodality, we learn piecewise mappings from observations to human poses. We first construct a tree on the pose manifold by applying affinity propagation clustering at the different levels of the hierarchy. Support vector machines classifiers are then trained to learn traversing the tree so that new examples/ observations can be classified to the clusters associated to the leaf nodes. Multi-valued Relevance Vector Machine (RVM) regressors are trained at each of the leaf nodes, so to learn local mappings from the observation to the pose space. We propose the use of a geodesic distance during clustering and describe a method for training multi-valued RVMs. The latter alleviates the need to train a separate RVM for each of the dimensions in the pose space. We validate the proposed method using the HumanEva dataset and show promising results.