Scale- and orientation-invariant keypoints in higher-dimensional data

Abstract

Description of keypoints, or local image features, is widely employed in computer vision. However, the most successful techniques do not extend immediately to more than two spatial dimensions. In this paper, we describe robust methods for extracting local orientations and gradient histograms from higher-dimensional data, using these techniques to develop a three-dimensional analogue of the popular Scale-Invariant Feature Transform (SIFT). We apply our algorithm to intra-patient registration of magnetic resonance (MR) images, with promising results. Our implementation will be released as open-source software.