Local Feature Analysis for real-time Action Recognition
Xuemei Xie, Wang Li, Jianan Li, Xun Xu, Kai Jin · 2018
The real-time Action recognition is necessary for enabling computer automatically recognize human action in real world video. However, the current architectures (e.g. Hidden two stream ConvNets) are relatively shallow and based on fully-connected structure, which cares about the full image. The network tends to fail if two videos share similar backgrounds. To address this issue, we take a deep look into the effectiveness of local feature for better action representation. In particular, we visualize the hidden two stream network as an example and observe that local feature indeed reduce the impact of background information. Finally, we give the experimental results to demonstrate how does the local feature affects the recognition performance. We verify the performance of our proposed network on the standard video dataset UCF101 and it achieves the recognition accuracy of 91.6%, achieving a 2.6% improvement over the state-of-the-art real-time approaches.