Feature-preserving mesh denoising via bilateral normal filtering

In this paper, we propose a feature-preserving mesh denoising algorithm which is effective, simple and easy to implement. The proposed method is a two-stage procedure with a bilateral surface normal filtering followed by integration of the normals for least squares error (LSE) vertex position updates. It is well-known that normal variations offer more intuitive geometric meaning than vertex position variations. A smooth surface can be described as one having smoothly varying normals whereas features such as edges and corners appear as discontinuities in the normals. Thus we cast feature-preserving mesh denoising as a robust surface normal estimation using bilateral filtering. Our definition of "intensity difference" used in the influence weighting function of the bilateral filter robustly prevents features such as sharp edges and corners from being washed out. We will demonstrate this capability by comparing the results from smoothing CAD-like models with other smoothing algorithms.