Hyperplane arrangements for the fast matching and classification of visual landmarks

Abstract

Many robotics and mechatronics systems rely on a fast analysis of visual landmarks. Recently, binary feature representations of the popular SIFT and SURF landmarks have been proposed that offer large speed improvements and low memory consumption at high accuracy. In this paper, we compare a binarisation based on median-centred hyperplanes to the dominating approach of random hyperplanes. We describe the algorithms in a joint taxonomy and show that the kernel for median-centred hyperplanes satiesfies Mercer's condition. Speed and accuracy are benchmarked in a registration and classification task. Both methods achieve the same dramatic speedup in kernel evaluation. But we show that median-centred hyperplanes are faster in binarisation, find better matches and generalise better over pose and individual variation in the classification.