Temporal Recursive Propagation Network for Action Recognition
Li Xu, Ming Zeng, Jinjun Wang · 2020 International Conference on Computer Engineering and Application (ICCEA) · 2020
Efficient spatio-temporal information modeling is the key to action recognition. Present state-of-the-art suffers from the trade-off of spatiotemporal information modeling capability and model complexity. In this paper, we propose a novel temporal recursive propagation network (TRP) which can efficiently encode and fusion spatiotemporal information. TRP Module can be inserted in existing 2D CNN architectures such as ResNet and MobileNet. Abundant experiments show that TRP enjoys the performance over 3D CNN at lower computational cost than 2D CNN. We evaluate the proposed TRP on the large action recognition benchmark dataset UCF-101. TRP outperforms the state-of-the-art methods on UCF-101 from scratch.