Action recognition via pose-based graph convolutional networks with intermediate dense supervision

Highlights：

• We propose a pose-based graph convolutional network (PGCN), which employs the graph convolutional module to model the spatiotemporal correlations among the pose-related features to produce a highly discriminative representation for human action recognition.

• We point out the laziness problem of the backbone CNN, and further propose a novel intermediate dense supervision (IDS) to solve this problem. It is simple and effective, without the need for extra parameters and computations.

• We evaluate our approach on three popular benchmarks for pose-based action recognition, where our approach achieves stateof-the-art performance on all of them.