Action Recognition Network Based on Temporal Spatial Temporal Mode

In recent years, it has become a trend to train an action recognition network with large-scale datasets, but under the condition of limited computing resources, the batch size cannot take a larger value, which will lead to the degradation of recognition performance. Based on this phenomenon, this paper proposes a brain-like convolution kernel module more suitable for small batch size training, which adopts the "Temporal - Spatial-Temporal" mode of thinking (TST). For small batch size, the proposed network using TST convolution module outperforms the baseline ST-GCN and the RA-GCN methods in the NTU-RGB+D dataset, showing its efficiency and feasibility.