Adaptive self-refining joint bilateral normal filtering for mesh denoising

The bilateral filter has been widely used for decades in various fields, including image processing and mesh denoising. Extending its capabilities, the joint bilateral filter was introduced to incorporate a guidance field, improving filtering quality by leveraging auxiliary information. While joint bilateral filtering guided by surface normals is highly effective for mesh denoising, current approaches rely on globally fixed Gaussian-kernel parameters, inherently limiting their ability to adapt to complex local geometry. In this paper, we propose a fully anisotropic, self-refining joint bilateral normal filtering framework. Our method overcomes global parameter limitations by dynamically estimating per-face spatial and range variances based on local geometric characteristics. Furthermore, we introduce an iterative refinement scheme that progressively updates these variance parameters as mesh normals are filtered, allowing the algorithm to self-adapt during the denoising process. Extensive experiments demonstrate that our approach significantly outperforms existing classical techniques, achieving an average error reduction of approximately 20% compared to baseline guided normal filtering. In addition, the proposed method proved to be effective against a recent optimization method and a tested deep learning approach, while maintaining computational efficiency without requiring offline training datasets.