Graph Convolutional Network-based Sports Skill-level Recognition via Deep Metric Learning

A Graph Convolutional Network (GCN)-based sports skill-level recognition method through Deep Metric Learning (DML) is proposed in this paper. For the realization of accurate skill-level recognition, it is necessary to derive feature representation, which can clearly discriminate skill levels on feature space, from motion data obtained from sports play-ers. Then our method carries out skill-level recognition using confidence-aware spatial-temporal attention graph convolutional network, which was trained by minimizing a cross entropy loss function and a center loss function. By minimizing these loss functions, our method enables the recognition using features that can discriminate their skill-levels on the feature space.