Real-time inference of 3D human poses by assembling local patches

This paper presents a novel patch-based approach for 3D human body pose estimation from a static silhouette image. Our approach uses a database which contains example patches extracted from different parts of images rendered using a 3D human body model in various poses. Each example patch is represented as a shape context histogram and contains pose parameters of the model that is used to render the image. At the estimation step, example patches in the database that have a similar shape context to patches extracted from the input silhouette image are rapidly retrieved using a modified locality-sensitive hashing algorithm. The pose parameters are then estimated by a probabilistic Hough voting scheme in a hierarchical manner. Our approach is flexible and needs a small number of examples since combinations of local patches can be used to identify previously unseen entire poses. Thus, it is not necessary for every possible pose to be stored in the database. This property significantly reduces computation time. Experiments have shown that our proposed method can handle a variety of unseen articulations and output accurate pose estimations in real time.