| A set of modular algorithms for efficiently finding SIFT correspondences in images or image archives is presented. The basic algorithm, called SIFT-HHM, exploits properties of the SIFT descriptor vector to find shortcuts to the most likely matches in two feature sets. SIFT-HHM converges approximately 15 times faster than a linear search, and, respectively, four and five times faster than PCA-SIFT and SURF at near-equivalent precision-recall performance.;A PCA-based binning algorithm that can be combined with SIFT-HHM is presented to address the content-based image retrieval problem. Our experiments show this combined approach to be preferable over current tree-based methods for a number of reasons. Most significantly, it will converge approximately three times faster than the current state ofthe art. Secondly, database build times are less than 10% of those for constructing a k-means tree. Finally, we note simplicity of storage, scalability, and suitability to distributed processing as incidental benefits. |
