Binary Keypoint Descriptor for Accelerated Matching

Abstract

Efficient extraction of key points from images is a hot topic in computer vision and forms many applications. We propose a kind of binary descriptor which is invariant to rotation, viewpoint change, blur change, brightness change, and JPEG compression. To best address the whole process, this paper covers key point detection, description and matching. Orientations of the interest points are estimated by the Haar-wavelet responses to achieve rotation invariance. Binary descriptors are computed by comparing the intensities of two points in the overlapping sampling pattern on image patches. At last, binary descriptors are matched by Hamming distance which can be done very fast on SSE instruction set of modern CPUs such as Core i7 processor. We use coarse-to-fine strategy to accelerate the matching of key point. In the experiment results, we will show that our descriptor is fast and robust.