Action Recognition from a Single Web Image Based on an Ensemble of Pose Experts

In this paper, we present a new method which estimates the pose of a human body and identifies its action from one single static image. This is a challenging task due to the high degrees of freedom of body poses and lack of any motion cues. Specifically, we build a pool of pose experts, each of which individually models a particular type of articulation for a group of human bodies with similar poses or semantics (actions). We investigate two ways to construct these pose experts and show that this method leads to improved pose estimation performance under difficult conditions. Furthermore, in contrast to previous wisdoms of combining the output of each pose expert for action recognition using such method as majority voting, we propose a flexible strategy which adaptively integrates them in a discriminative framework, allowing each pose expert to adjust their roles in action prediction according to their specificity when facing different action types. In particular, the spatial relationship between estimated part locations from each expert is encoded in a graph structure, capturing both the non-local and local spatial correlation of the body shape. Each graph is then treated as a separate group, on which an overall group sparse constraint is imposed to train the prediction model, with extra weight added according to the confidence of the corresponding expert. We show in our experiments on a challenging web data set with state of the art results that our method effectively improves the tolerance of our system to imperfect pose estimation.