3D point representation for pose estimation: Accelerated SIFT vs ORB

Abstract

Many novel local image descriptors (Random Ferns, ORB etc) are being proposed each year with claims of being as good as or superior to SIFT for representing point features. In this context we design a simple experimental framework to compare the performances of different descriptors for realtime recognition of 3D points in a given environment. We use this framework to show that robust descriptors like SIFT perform far better when compared to fast binary descriptors like ORB if matching process uses approximate nearest-neighbor search (ANNS) for acceleration. Such an analysis can be very useful for making appropriate choice from vast number of descriptors available in the literature.We further apply machine learning techniques to obtain better approximation of SIFT descriptor matching than ANNS. Though we could not improve its performance, our in-depth analysis of its root cause provides useful insights for guiding future exploration in this topic.