A Graph Convolutional Network with Ensemble Classifiers for 2D Skeleton-based Action Recognition

In this paper, we address the problem of action recognition from 2D skeletons produced by pose estimation algorithms. In practical applications, inaccurate joint information is inevitable due to self-occlusion and interaction with objects, which pose great challenges to 2D skeleton-based action recognition. To this end, we propose a graph convolutional network with ensemble classifiers to improve the robustness against noisy skeleton data. A global classifier learned from representations of the entire skeleton can capture the overall movement of the human body, while a set of part-based classifiers focuses on crucial local motion patterns. To further explore the correlations among joints within a part, we apply a self-attention mechanism to obtain context-aware part representations. Considering the existence of noisy joints, we develop a confidence-based weighting strategy to suppress classifiers based on inaccurate part representations. Experiments on the large-scale NTU-RGB+D dataset verify that the ensemble classifiers can boost the action recognition performance, especially on low-quality skeleton data with dense noise.