MVAvatar: Mesh-Guided Visibility-Aware 3D Avatar Reconstruction from Monocular Video

Abstract

Reconstructing high-fidelity 3D human avatars from monocular videos has attracted significant attention in computer vision. However, explicitly addressing self-occlusions and recovering unobserved body regions remain challenging, often resulting in noticeable artifacts. To resolve these limitations, we propose MVAvatar, a novel two-stage 3D avatar reconstruction framework. MVAvatar leverages mesh-based visibility analysis to spatially guide avatar optimization, introducing a visibility-aware strategy that decouples reconstruction constraints for observed regions from score distillation sampling-based generative guidance for occluded regions, achieving superior texture fidelity and stronger generalization to out-of-distribution driving poses.