Light-weight Enhanced Semantics-Guided Neural Networks for Skeleton-Based Human Action Recognition

In the skeleton-based human action recognition domain, the methods based on graph convolutional networks have had great success recently. However, most graph neural networks rely on large parameters, which is not easy to train and take up a large computational cost. In the above, a simple yet effective semantics-guided neural network (SGN) obtains with a few parameters and has achieved good results. However, the simple use of semantics is limited to the improvement of recognition rate. Moreover, using only one fixed temporal convolution kernel, which is not enough to extract the temporal details comprehensively. To this end, we propose an enhanced semantics-guided neural network (ESGN) in this paper. Some simple but effective strategies are applied to ESGN, such as semantic expansion, graph pooling methods, and regularization loss function, which do not significantly increase the parameter size but improve the accuracy on two large datasets than SGN. The proposed method with an order of magnitude smaller size than most previous papers is evaluated on the NTU60 and NTU120, the experimental results show that our method achieves the state-of-the-art performance.