Spatial Temporal Transpose Convolutional Networks for Action Recognition

In recent years, the use of skeleton data for human action recognition has received increasing attention in fields such as intelligent monitoring, human-object interaction, and video retrieval. Among them, Modeling Human Skeleton as Spatial-Temporal Graph Convolutional Network (ST-GCN) has made significant progress. However, methods based on Graph Convolution (GCN) generally face two issues: on one hand, the receptive fields of the spatial and temporal graphs are not flexible enough, limiting their feature representation capabilities. On the other hand, the model is overly parameterized, leading to high computational costs. In this paper, we first propose a spatial-temporal transpose convolution module to overcome the inflexibility of the spatial-temporal graph receptive field. Secondly, we introduce 1 *1 convolutions and depth-wise separable convolution layers into the network model, effectively reducing the computational complexity of the model. On the two subsets of the large-scale skeleton action recognition dataset NTU RGB-D, experimental results show that the proposed method achieved Top1 accuracy metrics of 84.6% and 92.5% respectively, while the model size was reduced by 13.48 times compared to the classic ST-GCN method, and the computational complexity was reduced by approximately 17.18 times.