Learning Silhouette Features for Control of Human Motion
Liu Ren, Gregory Shakhnarovich, Jessica K. Hodgins, Hanspeter Pfister, Paul A. Viola · 2004
features that are computed on the silhouette images [Viola and Jones 2001]. These features are computationally efficient and therefore suited to real-time applications. The best local features for estimating yaw and body configuration are selected from a very broad set of possible features based on a variant of the AdaBoost algorithm [Schapire and Singer 1999] (Figure 2). Synthetic data are used for the feature selection. The determination of the yaw orientation incorporates a fast search method called locality sensitive hashing. The determination of the body configuration (joint angles, root position, and orientation) relies on the temporal coherence in a domain-specific database of human motion. The database is pre-processed into an augmented motion graph that extends the motion graph [Lee et al. 2002] to handle scaling in velocity and two interacting characters. The search for the pose of the animated character is performed locally in the augmented motion graph in the region aroun