Region-based saliency estimation for 3D shape analysis and understanding

The detection of salient regions is an important pre-processing step for many 3D shape analysis and understanding tasks. This paper proposes a novel method for saliency detection in 3D free form shapes. Firstly, we smooth the surface normals by a bilateral filter. Such a method is capable of smoothing the surfaces and retaining the local details. Secondly, a novel method is proposed for the estimation of the saliency value of each vertex. To this end, two new features are defined: Retinex-based Importance Feature (RIF) and Relative Normal Distance (RND). They are based on the human visual perception characteristics and surface geometry respectively. Since the vertex based method cannot guarantee that the detected salient regions are semantically continuous and complete, we propose to refine such values based on surface patches. The detected saliency is finally used to guide the existing techniques for mesh simplification, interest point detection, and overlapping point cloud registration. The comparative studies based on real data from three publicly accessible databases show that the proposed method usually outperforms five selected state of the art ones both qualitatively and quantitatively for saliency detection and 3D shape analysis and understanding.

Neurocomputing, 197(July), 1–13. Elsevier