Improved Two-stream Network for Action Recognition in Complex Scenes
Yuxin Wang, Weibin Liu, Weiwei Xing · 2021
Compared with the great success of deep convolution network in still image visual recognition, the performance improvement of action recognition is limited by complex scenes such as complex background and fast motion of the object in the video. In this paper, we propose an improved two-stream network to solve the above problems. In the spatial stream network, we design a detection-before-recognition (DBR) strategy to maximize the proportion of effective information in the input image (video frame) and reduce the interference of complex background. In the temporal stream network, an attention mechanism is introduced, which could make the network focus on the object to be recognized and then to optimize the feature blur problem caused by the fast motion of the object. Finally, through the weighted linear combination of their prediction scores, the complex scene problem can be solved and the final recognition results can be improved. Experiments show that our method can get competitive performance on the UCF101 dataset.