Action-Scene Model For Recognizing Human Actions From Background In Realistic Videos

Using single information from person region fails to distinguish similar actions in realistic videos due to occlusions and variation of person. In this paper, we explore the problem of modeling action-scene context from the background regions of the realistic videos. The contextual cues of actions and scenes are formulated in a graphical model representation. A novel Action-Scene Model is proposed to mine the contextual cues with little prior knowledge. The proposed approach can infer actions from background regions directly and is a complement to the existing methods. In order to fuse the contextual cues effectively with other components, a context weight is introduced to measure the contributions of context based on the proposed model. We present experimental results on a realistic video dataset. The experiment results validate the effectiveness of Action-Scene Model in identifying the actions from background regions. And the learned contextual cues can achieve better performance than the existing methods especially for scene-dependent action categories.