Inverted indexing in image fragment retrieval using huge keypoint-based vocabularies

Inverted indexing over a huge keypoint-based vocabulary is proposed. For each keypoint, a set of words representing photometric and geometric properties of the keypoint context is built using triplets of neighboring keypoints. The method is intended primarily for the retrieval of images containing similar fragments (partial near-duplicates). With a size of vocabulary exceeding by several orders the total number of web images, inverted files are very small. Thus, even in very large databases the number of images retrieved for a given query is expected to be small (and precision high). Moreover, we claim (based on limited-scale experimental results) that the numbers of retrieved images do not grow proportionally to the database size.