OAHPE: occlusion-aware hybrid routing for efficient and robust 3D human pose estimation in mixed-occlusion videos.

Abstract

Attention-mechanism-based methods for video-based 3D human pose estimation (HPE) have demonstrated strong performance in spatio-temporal modeling. However, the quadratic complexity introduced by attention leads to substantial computational and memory overhead during long-sequence inference, thereby hindering lightweight deployment. Inspired by the long-range modeling capability of the state space model (SSM) with linear complexity, we propose the Occlusion-Aware Hybrid Pose Estimation (OAHPE) model to address the dual challenges of high uncertainty in occluded frames and high redundancy in visible frames during 3D human pose estimation from video. This model employs a differentiated dual-path modeling strategy. For occluded segments, we integrate a multi-scale dilated convolutions with linear-complexity-based Mamba module, and employ a dual local-enhancement mechanism to capture fine-grained joint structures and long-range inter-joint dependencies at the full frame rate. For visible segments, we design an online selection mechanism to reduce temporal redundancy, and introduce a distillation-based temporal recovery objective to enforce dynamic consistency. Finally, predictions from the two paths are backfilled, aligned, and fused at the original sequence resolution. Comprehensive quantitative and qualitative evaluations on the Human3.6M and MPI-INF-3DHP benchmarks demonstrate that OAHPE achieves mean per-joint position errors of 43.5 mm and 28.5 mm, respectively, while reducing the model size and training memory footprint to 2.6 M parameters and 12.5 GB. These results indicate that the proposed method attains superior performance with substantially lower parameter and memory costs, achieving a favorable trade-off between pose estimation accuracy and computational efficiency.