The Hand Reconstruction Method Based on Diffusion Model

Le résumé fourni par la source

Hand mesh reconstruction is critical for applications in virtual reality, human-computer interaction, and biomedical engineering. Existing methods struggle with generalization under dynamic poses, occlusions, and cross-domain scenarios (e.g., surgical environments), while traditional approaches lack robustness to noise and deep learning techniques demand heavy computational resources. This paper introduces a novel framework integrating frequency filtering and diffusion models to address these limitations. First, frequency-domain low-pass filtering enhances input quality by suppressing noise (e.g., motion blur) and emphasizing anatomical features (e.g., joint contours). Second, a diffusion model progressively refines 3D hand vertices and joints through iterative denoising, resolving depth ambiguity and occlusion challenges. A hierarchical Transformer further strengthens robustness by fusing diffusion priors with multi-scale image features. Evaluated on the InterHand2.6M benchmark, our method achieves a 13.4% reduction in median hand surface error (0.142 cm to 0.123 cm) and a 13.2% decrease in keypoint error (1.52 cm to 1.32 cm), with inference speed accelerated by 30% (20 ms per frame). Cross-domain tests in surgical settings demonstrate stable performance under severe noise and occlusions. Key contributions include: (1) a synergistic framework combining frequency filtering and diffusion models for high-precision reconstruction, (2) a resource-efficient hierarchical fusion strategy reducing dependency on labeled data, and (3) comprehensive validation across diverse scenarios. Future work will focus on lightweight diffusion architectures and joint frequency-spatial optimization to expand practical applicability.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.