BoxNet: Fine-grained Action Classification in Boxing

The skeleton data is widely used in general action recognition since it is robustly distinguish the action operator and complex background. When it comes to more fine-grained scenario like those in sports competitions, such characteristic of skeleton data could further enlarge differences of actions among varied classes and reduce negative effects from background to achieve better recognition result. However, the problem of coarse video record which severely deteriorate the validity of neural network is still to be solved. In this work, we present an improved graph convolutional network(GCN) methods of fine-grained technical action classification based on previous works, where early fusion and parameterized adjacency matrix are also adopted. An interpolation method is firstly applied in action classification to elevate the performance of networks with coarse input. Moreover, a new dataset about boxing established for fine-grained action classification task. Our model is verified on NTU-RGB+D and tested on established boxing dataset, and exceeds common GCN methods.