A quantitative evaluation of comprehensive 3D local descriptors generated with spatial and geometrical features

摘要

Designing a feature descriptor with high descriptiveness and strong robustness is an important task in 3D computer vision. Both encoding spatial and geometrical information are critical for a 3D local descriptor. In this paper, we comprehensively explore the performance of different methods for encoding spatial information with both the polar coordinate system and the Cartesian coordinate system, and also investigates the performance of combining these spatial methods with one effective geometrical attribute (i.e., normal deviation angle). Eventually, 26 descriptors are generated. Among these descriptors, eleven are generated with the polar coordinate system; fourteen are generated with the Cartesian coordinate system; one is constructed of only encoding the geometrical information. Extensive experiments are conducted for evaluating the 26 descriptors on four benchmark datasets in terms of descriptiveness, robustness, compactness, efficiency, and application performance. Based on the experimental results, the traits, merits and demerits of all the 26 descriptors are summarized. The results present that the descriptors generated by different combination of spatial attributes and geometrical attribute have significant influence to their performance. The results also show that some descriptors designed in this paper have superior overall performance compared with the state-of-the-art ones.