Mesh saliency with global rarity

Abstract

Reliable estimation of visual saliency is helpful to guide many computer graphics tasks including shape matching, simplification, segmentation, etc. Inspired by basic principles induced by psychophysics studies, we propose a novel approach for computing saliency for 3D mesh surface considering both local contrast and global rarity. First, a multi-scale local shape descriptor is introduced to capture local geometric features with various regions, which is rotationally invariant. Then, we present an efficient patch-based local contrast method based on the multi-scale local descriptor. The global rarity is defined by its specialty to all other vertices. To be more efficient, we compute it on clusters first and interpolate on vertices later. Finally, our mesh saliency is obtained by the linear combination of the local contrast and the global rarity. Our method is efficient, robust, and yields mesh saliency that agrees with human perception. The algorithm is tested on many models and outperformed previous works. We also demonstrated the benefits of our algorithm in some geometry processing applications.