On creating low dimensional 3D feature descriptors with PCA

摘要

With the availability of commodity depth sensors and depth sensing capabilities on mobile devices, there is a need to develop memory efficient and computationally cheap applications. 3D feature descriptor based keypoint matching is the very first step in various computer vision and robotic applications, and is also one of the most memory and computationally demanding steps. In this work, we apply Principal Component Analysis (PCA) to reduce the dimensionality of 3D feature descriptors and show that they retain their descriptiveness, while drastically decreasing their memory and computational requirements. We apply this scalable PCA based dimensionality reduction technique on three state-of-the-art 3D feature descriptors and quantitatively evaluate their performance with varying dimensionality. Experimental results on publicly available Bologna Kinect dataset show that the proposed method reduces the dimensionality of SHOT, RoPS and FPFH feature descriptors from 352 to 50, 135 to 30 and 33 to 15, respectively, while offering more than 90% of the keypoint matching performance.