Multi-scale mesh saliency based on low-rank and sparse analysis in shape feature space

Abstract

This paper advocates a novel multi-scale mesh saliency method using the powerful low-rank and sparse analysis in shape feature space. The technical core of our approach is a new shape descriptor that embraces both local geometry information and global structure information in an integrated way. Our shape descriptor is organized in a layered and nested structure, enabling both multi-scale and multi-level functionalities. Upon devising our novel shape descriptor, the remaining challenge is to accurately capture sub-region (or sub-part) saliency from 3D geometric models. Towards this goal, we exploit our novel shape descriptor to define local-to-global shape context in a vertex-wise fashion and concatenate all the shape contexts to form a feature space, which encodes both local geometry feature and global structure feature. It then paves the way for us to employ the powerful low-rank and sparse analysis in the feature space, because the low-rank components emphasize much more on stronger patch/part similarities, and the sparse components correspond to their differences. By focusing on the sparse components, we develop a versatile, structure-sensitive saliency detection framework, which can distinguish local geometry saliency and global structure saliency in various 3D geometric models. Our extensive experiments have exhibited many attractive properties of our novel shape descriptor, including: being suitable for perception-driven analysis, being structure-sensitive, multi-scale, discriminative, and effectively capturing the intrinsic characteristic of the underlying geometry.