SBL Mesh Filter: A Fast Separable Approximation of Bilateral Mesh Filtering

Abstract

Bilateral mesh filtering is a simple and powerful feature-preserving filtering operator which allows to smooth or remove noise from surface meshes while preserving important features in a non-iterative way. However, to be effective, such a filter may require quite a large support size, inducing slow processing when applied on high resolution meshes such as the ones produced by automatic 3D acquisition devices. In this paper, we propose a separable approximation of bilateral mesh filtering based on a local decomposition of the bi-dimensional filter into a product of two one-dimensional ones. In particular, we show that this approximation leads to piecewise smooth surfaces which are very close to the ones produced by the exact filter, using only a fraction of the usual required time. Compared to bilateral image filtering, the main problem here is to find meaningful directions at every point to orient the two one-dimensional filters. Our solution exploits the minimum and maximum curvature directions at each point and demonstrates a significant speed-up on meshes ranging from thousands to millions of elements, enabling feature-preserving filtering with large support size in a variety of practical scenarii. Our approach is simple, easy to implement and orthogonal to other kinds of optimizations such as higher dimensional clustering using a bilateral grid or a Gaussian kd-tree and can therefore be combined to them to reach even higher performance.