Abstract | ||
---|---|---|
Video action recognition is a vital area of computer vision. By adding temporal dimension into convolution structure, 3D convolution neural network owns the capacity to extract spatio-temporal features from videos. However, due to computing constraints, it is hard to input the whole video into the convolution network at one time, resulting in a limited temporal receptive field of the network. To address this issue, we propose a novel 3D temporal dilation convolution (3D-TDC) framework, to extract spatio-temporal features of actions from videos. First, we deploy the 3D temporal dilation convolution as the shallow temporal compression layer, enabling an effective capture of spatio-temporal information in a larger time domain with the reduced computational load. Then, an action recognition framework is constructed by integrating two networks with different temporal receptive fields to balance the long-short time difference. We conduct extensive experiments on three widely-used public datasets (UCF-101, HMDB-51, and Kinetics-400) for performance evaluation, and the experimental results demonstrate the effectiveness of our proposed framework in video action recognition with low computational load. |
Year | DOI | Venue |
---|---|---|
2021 | 10.1016/j.neucom.2021.03.120 | Neurocomputing |
Keywords | DocType | Volume |
3D convolution,Temporal dilation,Action recognition,Temporal compression | Journal | 450 |
ISSN | Citations | PageRank |
0925-2312 | 0 | 0.34 |
References | Authors | |
0 | 4 |
Name | Order | Citations | PageRank |
---|---|---|---|
Yue Ming | 1 | 0 | 0.34 |
Fan Feng | 2 | 2 | 1.52 |
Chao Li | 3 | 0 | 0.34 |
Jing-Hao Xue | 4 | 15 | 10.05 |