Fine-Grained Video Action Recognition via Mutually Reinforced Graph Convolution and Cross-Fiber Interaction

Abstract

Currently, sophisticated video action recognition remains a highly challenging task, primarily due to the intricate interactions between local motion primitives and their broader temporal context, relationships that existing methods have not fully elucidated. This paper explores a framework that attempts to integrate two complementary approaches: one emphasizing the extraction of spatiotemporal “tubular structures” as atomic action units through mutually reinforcing learning, and the other promoting cross-fiber collaborative enhancement across hierarchical feature dimensions. Our preliminary experiments on benchmark datasets show that the performance improvement compared to the baseline configuration is modest, but the improvement appears to be quite sensitive to input resolution and temporal sampling strategies, suggesting some unresolved dependencies worthy of further investigation.