MFA-Net: Motion Field Adaptive Network for Skeleton-Based Action Recognition

In recent years, skeleton-based action recognition has made significant progress but still faces several pressing challenges. Notably, the motion field ranges of different actions vary significantly, and existing methods struggle to simultaneously capture macromorphology and microdetail features of actions. This limitation impedes the accurate modeling of action-specific patterns. To address this issue, we propose a novel Motion Field Adaptive Network (MFA-Net). This network employs a parallel dual-stream architecture, comprising a Macromorphology Field Perception Stream (MmFPS) and a Microdetail Field Perception Stream (MdFPS) respectively. Specifically, MmFPS introduces a spatio-temporal collaborative perception mechanism to model global dependencies in macroscopic posture evolution, capturing the complete semantics and dynamic trends of actions. Meanwhile, MdFPS contains layer-wise modulated spatial graph convolution and multi-scale temporal field convolution, enhancing the representation of local spatio-temporal features. Additionally, we present a Motion Field Adaptive Fusion (MFAF) strategy to fully leverage the advantages of both streams while allowing the model to flexibly adjust the relative weights of macromorphology and microdetail features based on the motion field ranges of actions, thus achieving more precise action recognition. Experimental results on three representative benchmark datasets demonstrate that the proposed method achieves state-of-the-art performance.