A progressive decoding strategy for action recognition

In recent years, significant progress has been made in modeling temporal sequences and spatial structures skeleton-based human action recognition. However, existing methods rely on explicit of the inherent structure body, which may result reduced joint saliency poor interpretability due to sparsity skeleton data relative smoothness convolutions. This paper proposes a feature matching method based progressive decoding strategy. As movement is chain process, strategy progressively decodes pose features from center periphery, using multi-level graph filters obtain multi-frequency hierarchical features. Then adaptive convolution kernels are constructed match local similarities between Self-similarity query set mutual similarity support samples calculated analyze entire posture similar skeletal distinguished according node spectrum differentiate different categories. Through experimental verification two public sets, proposed better recognition accuracy generalization small sample behavior. The experiments show that outperforms NTU RGB + D Human36M Dataset.