ST-CFNet: Spatio-Temporal Cross-Feature Fusion Networks for 3D Human Pose Estimation

Despite the remarkable progress in 3D human pose estimation, existing methods face dual challenges of insufficient spatiotemporal modeling and inefficient cross-modal interaction in complex motion and occlusion scenarios. To address this, we propose a novel network named ST-CFNet, which leverages a dual-stream interaction framework to separately capture the spatial structural features and temporal dynamic patterns of the human body. Meanwhile, we design an Adaptive Cross-feature Fusion Module (ACFM), which constructs a decomposable channel interaction matrix to adaptively enhance spatiotemporal features and achieve cross-modal fusion, while suppressing irrelevant information. Compared to traditional parallel architectures, ST-CFNet enables progressive fusion of spatiotemporal cues during the hierarchical feature extraction process. Experimental results show that the model achieves significant performance improvements on benchmark datasets, such as Human3.6 M and MPI-INF-3DHP, while reducing estimation error.