A Spatio-Temporal Motion Network for Action Recognition Based on Spatial Attention

Temporal modeling is the key for action recognition in videos, but traditional 2D CNNs do not capture temporal relationships well. 3D CNNs can achieve good performance, but are computationally intensive and not well practiced on existing devices. Based on these problems, we design a generic and effective module called spatio-temporal motion network (SMNet). SMNet maintains the complexity of 2D and reduces the computational effort of the algorithm while achieving performance comparable to 3D CNNs. SMNet contains a spatio-temporal excitation module (SE) and a motion excitation module (ME). The SE module uses group convolution to fuse temporal information to reduce the number of parameters in the network, and uses spatial attention to extract spatial information. The ME module uses the difference between adjacent frames to extract feature-level motion patterns between adjacent frames, which can effectively encode motion features and help identify actions efficiently. We use ResNet-50 as the backbone network and insert SMNet into the residual blocks to form a simple and effective action network. The experiment results on three datasets, namely Something-Something V1, Something-Something V2, and Kinetics-400, show that it out performs state-of-the-arts motion recognition networks.