Mesh Denoising Transformer

Abstract

Mesh denoising, aimed at removing noise from input meshes while preserving their feature structures, is a practical yet challenging task. Despite the remarkable progress in learning-based mesh denoising methods in recent years, their network designs still suffer from two major limitations: a dependence on single-modal geometric representations, which fall short of capturing the multifaceted attributes of meshes, and a lack of effective global feature aggregation, which limits their ability to model the global mesh structure. To tackle these issues, we propose SurfaceFormer, a novel Transformer-based mesh denoising framework. Our first contribution is the development of a new representation known as Local Surface Descriptor, which is crafted by establishing polar coordinate systems on each mesh face, followed by sampling points from adjacent surfaces using geodesics. The normals of the sampled points are arranged into image-like 2D patches to capture local geometric details, whereas the poles and vertex coordinates are consolidated into a point cloud to represent spatial information. This representation alleviates the challenges posed by the irregular and non-Euclidean characteristics of mesh data, facilitating a smooth integration with the Transformer architecture. Next, we propose a dual-stream structure consisting of a Geometric Encoder branch and a Spatial Encoder branch, which jointly encode local geometry details and spatial information to fully explore multimodal information for mesh denoising. A subsequent Denoising Transformer module receives the multimodal information and achieves efficient global feature aggregation through self-attention operators. Our experimental evaluations demonstrate that this novel approach outperforms existing state-of-the-art methods in both objective and subjective assessments, representing a significant advance in mesh denoising.