Evaluation of Feature Detectors, Descriptors and Match Filtering Approaches for Historic Repeat Photography

摘要

This work analyzes the suitability of classic feature detectors and descriptors as well as different match filters for matching historical to modern images. A variety of prominent detector and descriptor combinations are evaluated on a new dataset, that is composed of repeat photographs of various scenes exposed to tremendous change across years. Results show that a dense keypoint sampling is more effective than classic feature detection, while several descriptors achieve comparable performances. Yet, more important is an adequate match filtering approach that identifies correct correspondences despite their large descriptor distances. We show that the standard ratio test is unsuitable for matching historic to modern image pairs, while other filters based on keypoint geometry boost performance. Finally, we establish a complete pipeline including detectors, descriptors and match filtering methods. Our pipeline shows high performance on challenging image pairs beyond repeat photography, as our tests on another dataset show.