Reproduce “Mesh Denoising Transformer” on OpenResearch

About this paper

Wenbo Zhao, Xianming Liu, Deming Zhai, Junjun Jiang, Xiangyang Ji

This paper presents SurfaceFormer, a pioneering Transformer-based framework for mesh denoising that introduces a novel multimodal Local Surface Descriptor to transform irregular mesh data into a format suitable for global context modeling. The model consistently achieved superior quantitative and qualitative results compared to prior state-of-the-art methods, accurately preserving fine features while effectively removing noise from various 3D mesh types.

Key insights

• Irregular mesh data can be effectively represented in a multimodal format (Local Surface Descriptor - LSD) that captures both local geometric details and global spatial information, making it compatible with Transformer architectures.
• A dual-stream encoder architecture, combining a lightweight ResNet for geometric features and a DGCNN for spatial features, efficiently extracts and processes distinct multimodal data from meshes.
• Integrating Transformer's self-attention mechanism enables efficient global feature aggregation and long-range dependency modeling, overcoming the localized receptive field limitations of prior CNN and GNN methods in mesh denoising.

Summary by alphaXiv.