PHOTO AND VIDEO CLUSTERING BY USE OF COMMUNITY DETECTION METHODS ON HYBRID SIMILARITY GRAPHS.

摘要

The invention pertains to information retrieval applications over large collections of photos and videos, where users face the problem of information overload. Photo and video clustering into groups of items that refer to the same topic or object considerably alleviates the above problem. The photo and video clustering method tackles the problem of clustering by use of a community detection method on hybrid similarity graphs. The hybrid similarity graph constitutes a data structure for the representation of different types of similarity between photos and videos, e.g. visual similarity or similarity based on the text metadata. The community detection method extracts from the hybrid similarity graph, subgraphs with higher connectivity between their members compared to the rest of the graph. Subgraphs with this characteristic form the resulting clusters. The invention is applicable to problems that require the review and navigation over very large amounts of photos and videos that are published from users in social networking applications (e.g. Facebook, Flickr, YouTube).