Action-HSMR: Sequence-based 3D Human Pose and Mesh Estimation with Temporal Consistency

The accurate estimation of multi-dimensional human pose recognition relies heavily on information about posture and movement. Specifically, estimating 3D human pose and mesh requires not only anatomically plausible reconstructions but also robustness to challenging scenarios. Existing anatomically constrained methods often process frames independently, leading to implausible predictions and temporally inconsistent results. To address this, we propose Action-HSMR, a sequence-based framework that leverages short-term temporal context from consecutive frames to correct potentially implausible SKEL parameter predictions. Specifically, sliding windows of three frames are serialized into structured spatial-temporal tokens, encoded via Vision Transformer (ViT), and fused using a self-attention temporal feature fusion module. Extensive experiments on 3D benchmark datasets HMR2.0, demonstrate state-of-the-art performance, particularly under extreme poses.