Local and Global Spatial-Temporal Transformer for skeleton-based action recognition

Download PDF
Open Webpage

Ruyi Liu, Yu Chen, Feiyu Gai, Yi Liu, Qiguang Miao, Shuai Wu

Published: 2025, Last Modified: 06 Jan 2026Neurocomputing 2025EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•This architecture allows the model to capture local and global features between joints more effectively and realize the full fusion of different modalities.•By adding physical topology constraints of human skeleton and integrating multi-scale temporal convolution and transformer, the model can learn more abundant information in spatial and temporal dimension better.•A new joint-bone cross fusion transformer(JBC-Former) is proposed, which dynamically generates complementary features based on the original input.