ResearchPod Summary
Human Mesh Recovery (HMR) has traditionally focused on single-person reconstruction or small, uncrowded scenes. These approaches often fail in real-world applications like autonomous driving or large-scale surveillance, where severe inter-person occlusions and high crowd density are common. This paper addresses the lack of suitable benchmarks and models for these complex, large-scale environments.
The authors introduce MVMP-HMR, a large-scale synthetic dataset featuring 15 complex scenes, up to 50 camera views, and 30 interacting persons per scene. To tackle the recovery task, they propose the MVMP-HMR model, which operates in three stages:
To further improve robustness against occlusions, the authors introduce two novel loss functions: an orientation loss to supervise the human mesh's global orientation and a 3D joint density loss to supervise 3D joint positions via density maps rather than direct coordinate regression.
Experiments demonstrate that existing state-of-the-art HMR methods struggle significantly on the MVMP-HMR benchmark due to the high degree of occlusion and scene scale. The proposed MVMP-HMR model consistently outperforms these prior methods, showing that fusing multiview features into a unified 3D volume, combined with specialized losses, effectively mitigates ambiguity in crowded, real-world-like scenarios.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.