A novel SDASS descriptor for fully encoding the information of a 3D local surface

Abstract

Local feature description is a fundamental yet challenging task in the 3D computer vision. This paper proposes a novel descriptor, named the statistic of deviation angles on subdivided space (SDASS), for encoding geometrical and spatial information of a local surface based on a local reference axis (LRA). Because the surface normal is vulnerable to various common nuisances, we propose a robust geometrical attribute, called the local minimum axis (LMA), that replaces the normal to generate the deviation angle between LMA and LRA in our SDASS descriptor. To encode spatial information, we use two spatial features to fully encode the spatial information on a local surface based on an LRA that can achieve higher overall repeatability than the local reference frame (LRF). Furthermore, an improved LRA is proposed for increasing the robustness of our SDASS to noise and varying mesh resolutions. The performance of our SDASS descriptor is rigorously tested on four popular datasets and two modified datasets. The results show that our SDASS has high descriptiveness and strong robustness, and is obviously superior to the existing algorithms. Finally, the proposed SDASS is applied to 3D registration. The accurate results further confirm the effectiveness of our SDASS method.