Near-duplicate retrieval: A benchmark study of modified SIFT descriptors

Abstract

Local feature detectors and descriptors are widely used for image near-duplicate retrieval tasks. However, most studies and evaluations published so far focused on increasing retrieval accuracy by improving descriptor properties and similarity measures. There has been almost no comparisons considering the modification of the descriptors and the impact on accuracy and performance, which is especially of interest for interactive retrieval systems that require fast system responses. Therefore, we evaluate in this paper accuracy and performance of variations of SIFT descriptors (reduced SIFT versions, RC-SIFT−64D, the original SIFT−128D) and SURF−64D in two cases: Firstly, using benchmarks of various sizes. Secondly, using one particular benchmark but extracting varying amounts of descriptors. Another aspect that has been almost neglected in previous benchmarks is the combination of different affine transformations in near-duplicate images. A problem that many realworld systems have to face. Therefore, we provide in addition results of a comparative performance analysis using benchmarks generated by combining several image affine transformations.