Adaptive similarity search in large databases - application to image/video copy detection

This paper addresses the problem of similarity searches within large databases of multimedia documents. It proposes an adaptive and parameter-free method for automatic decision making during the final step of the similarity search. This method computes the minimum number of features that have to be shared by two documents to be considered as similar. It is based on a probabilistic approach that uses an a contrario modelling of the similarity. The resulting decision threshold is well-adapted to the number of features in the database, the number of features in the query document and also their rareness. The method is applied to image and video copy detection, for which the features are the image local descriptors. Experiments show the effectiveness and efficiency of the proposed method.