Spatially related image mining on very large image collections

Collections of images of ever growing sizes are becoming common. Structuring and browsing large image databases is a challenging problem. Access to images based on the 3D acquisition location or on the spatial overlap of the scenes they depict is intuitive and has high user acceptability. Commonly, the sets of relevant spatially related images are obtained using manual annotations. A method for discovering spatial overlaps using image content only via image retrieval techniques was proposed. The objective of the proposed approach is to provide a randomized data mining method for finding clusters of images with spatial overlap. Instead of trying to match each image, in turn, the method relies on the min- Hash algorithm for fast detection of random pairs of images with spatial overlap, the so-called cluster seeds. The seeds are then used as visual queries and clusters are obtained as transitive closures of sets of partially overlapping images that include the seed. This approach shows that the probability of finding a seed for an image cluster rapidly increases with the size of the cluster and approaches one fast. For practical database sizes, the running time of the seed generation process is close to linear in the size of the database. The cluster completion process requires a number of visual queries proportional to the number of images in all clusters. The proposed method discovers the spatially related images in large-scale image databases.