Fine-grained action recognition using multi-view attentions

Inflated 3D ConvNet (I3D) utilizes 3D convolution to enrich semantic information of features, forming a strong baseline for human action recognition. However, 3D convolution extracts features by mixing spatial, temporal and cross-channel information together, lacking the ability to emphasize meaningful features along specific dimensions, especially for the cross-channel information, which is, however, of crucial importance in recognizing fine-grained actions. In this paper, we propose a novel multi-view attention mechanism, named channel–spatial–temporal attention (CSTA) block, to guide the network to pay more attention to the clues useful for fine-grained action recognition. Specifically, CSTA consists of three branches: channel–spatial branch, channel–temporal branch and spatial–temporal branch. By directly plugging these branches into I3D, we further explore the impact of location information as well as the number of blocks in terms of recognition accuracy. We also examine two different strategies for designing a mixture of multiple CSTA blocks. Extensive experiments demonstrate the effectiveness of our CSTA. Namely, while using only RGB frames to train the network, I3D equipped with CSTA (I3D–CSTA) achieves accuracies of 95.76% and 73.97% on UCF101 and HMDB51, respectively. These results are indeed comparable with the results produced by the methods using both RGB frames and optical flow. Even more, with the assistance of optical flow, the recognition accuracies of CSTA–I3D rise to 98.2% on UCF101 and 82.9% on HMDB51, outperforming many state-of-the-art methods.