Feature-preserving mesh denoising via attenuated bilateral normal filtering and quadrics

Abstract

In this paper, we propose a two phase feature-preserving mesh denoising algorithm. The first phase consists of modified bilateral filtering applied on field of face normals. The range filtering is affected by gradual attenuation of the standard deviation of normal differences along with subsequent iteration. We also provide a method for automatic estimation of the bilateral filter parameters without user interaction. Filtering process is followed by reconstruction of the mesh vertex positions from a given field of face normals. Our method is based on geometric fact that every vertex should lie within all tangent planes that locally support one of its neighbouring triangles. We use quadrics to encode squared distances to the planes defined by filtered normals and centroids of the faces from 1-ring neighbourhood of considered vertex. New vertex positions are identified by minimizing quadric error metric. Our tests have shown that this iteration scheme converges faster than gradient descent algorithm. Sharp features are well preserved which will be documented on several CAD-like models.