A Deep Model Combining Structural Features and Context Cues for Action Recognition in Static Images

摘要

In this paper, we present a deep model for the task of action recognition in static images, which combines body structural information and context cues to build a more accurate classifier. Moreover, to construct more semantic and robust body structural features, we propose a new body descriptor, named limb angle discriptor (LAD), which uses the relative angles between the limbs in 2D skeleton. We evaluate our method on the PASCAL VOC 2012 Action dataset and compare it with the published results. The result shows that our method achieves 90.6% mean AP, outperforming the previous state-of-art approaches in the field.