MoPO: Incorporating Motion Prior for Occluded Human Mesh Recovery

Researchers introduce MoPO, a novel method for recovering human mesh models from occluded images by leveraging motion prediction from pose sequences. The approach combines spatial-temporal occlusion detection with lightweight motion prediction to estimate hidden body parts, achieving state-of-the-art results on occlusion benchmarks while reducing temporal inconsistencies.

MoPO addresses a critical limitation in human mesh recovery systems: their inability to handle occlusions with accuracy and temporal consistency. Traditional approaches rely primarily on image features, which become unreliable when body parts are hidden. The breakthrough here is conceptual—recognizing that historical pose sequences contain stronger predictive signals for occluded regions than degraded visual information alone. This insight reshapes how researchers approach the occlusion problem by treating it as a motion completion challenge rather than purely a visual reconstruction challenge.

The technical implementation combines two complementary mechanisms. A spatial-temporal occlusion detector identifies which joints lack reliable visual information, while a lightweight motion predictor infers plausible joint positions from historical poses. These predictions then fuse with image features for shape and pose estimation, with inverse kinematics providing final refinement. This architecture demonstrates how multi-modal reasoning—combining temporal dynamics with spatial vision—outperforms single-modality approaches.

The significance extends beyond academic benchmarks. Robust human mesh recovery under occlusion enables more reliable applications in sports analytics, motion capture, augmented reality, and surveillance systems where people are frequently partially hidden. The emphasis on temporal consistency addresses real-world motion jitter problems that plagued earlier systems. Performance improvements on both occlusion-specific and standard benchmarks suggest the method generalizes well rather than overfitting to specialized scenarios.

Future directions likely involve extending this motion-prior approach to more complex scenarios: multiple occluded people, extreme occlusion percentages, and diverse motion types beyond typical walking patterns. The lightweight design suggests practical deployment potential in resource-constrained environments.

• →Motion prediction from historical poses provides more reliable signals for occluded body parts than image features alone
• →MoPO combines spatial-temporal occlusion detection with motion prediction to achieve state-of-the-art results on occlusion benchmarks
• →The method reduces temporal jitter while improving pose accuracy, addressing key failure modes of prior human mesh recovery systems
• →Inverse kinematics refinement using predicted motion priors ensures anatomically consistent pose estimates
• →Multi-modal fusion of temporal dynamics and visual features outperforms single-modality approaches for handling occlusions