SasMamba: A Lightweight Structure-Aware Stride State Space Model for 3D Human Pose Estimation

On November 13, 2025, the SasMamba model was unveiled, representing a notable advancement in 3D human pose estimation. Built on the Mamba architecture and State Space Models (SSMs), SasMamba employs a structure-aware spatiotemporal convolution combined with a stride-based scan strategy. This innovative approach overcomes the limitations of traditional SSM methods, which often flatten 2D pose sequences into temporal sequences, thereby disrupting the spatial structure of human poses. By maintaining the integrity of spatial relationships, SasMamba captures complex pose dependencies more effectively. The model not only retains linear computational complexity but also achieves competitive performance in 3D pose estimation with significantly fewer parameters compared to existing hybrid models. This development is crucial for applications in AI and computer vision, as it enhances the efficiency and accuracy of pose estimation technologies.