Learning Hierarchical Context for Action Recognition in Still Images

Recognizing actions from still images is challenging due to the lack of movement information. Most of the existing works intend to characterize actions by the interaction context between each pair of features extracted from local image regions, which often fails to capture the complex structures in actions. Different from these works, in this paper we propose to divide the human body into a set of increasingly finer body parts, forming our hierarchical composition. To model the interaction patterns among these body parts, we further develop the hierarchical propagation network. By propagating information bottom-up in the composition, our model efficiently mines the interaction among local body parts, and integrates those discriminative context cues hierarchically into a compact action representation. Our experiments on the HICO and VOC 2012 Action datasets demonstrate the efficiency of our method for characterizing static actions.