Learned 3D mesh saliency from multi-scale spiral patch features

Mesh saliency, the process of identifying visually significant regions in three-dimensional (3D) meshes, plays a crucial role in computer graphics, with applications in tasks such as denoising and simplification. In this paper, we introduce a 3D mesh saliency measure designed to detect sharp geometric features. Our approach constructs a local normal-based feature for each vertex using a spiral path within a k-hop neighborhood. First, we compute geometric-based saliency as the mean local alignment of these spiral features. Next, spectral-based saliency is derived from the total variation of the structure tensor at each vertex, with gradients determined by spiral feature alignments. The final saliency measure is defined as vertex roughness weighted by the combined geometric and spectral saliencies. To extend this single-scale saliency to a multi-scale approach, we progressively decimate the mesh and integrate the results using max-pooling. In addition, rather than relying on fixed aggregation, we propose learning saliency as a regression problem using multi-scale roughness, geometric, and spectral saliencies. Our method is evaluated against state-of-the-art techniques across multiple datasets and demonstrates competitive performance.