Exploiting spatio‐temporal knowledge for video action recognition

Action recognition has been a popular area of computer vision research in recent years. The goal this task is to recognise human actions video frames. Most existing methods often depend on the visual features and their relationships inside videos. extracted only represent information current itself cannot general knowledge particular beyond video. Thus, there are some deviations these features, performance still requires improvement. In sudy, we present novel spatio-temporal module (STKM) endow with commonsense knowledge. To end, first collect hybrid external from universal fields, which contains both semantic information. Then graph convolution networks (GCN) used aggregate GCNs involve (i) spatial capture relations (ii) temporal serial occurrence among actions. By integrating can get better results. Experiments AVA, UCF101-24 JHMDB datasets show robustness generalisation ability STKM. results report new state-of-the-art 32.0 mAP AVA v2.1. On datasets, our method also improves by 1.5 AP 2.6 AP, respectively, over baseline method.