Unsupervised similarity learning through Cartesian product of ranking references

Despite the consistent advances in visual features and other Multimedia Information Retrieval (MIR) techniques, measuring the similarity among multimedia objects is still a challenging task for an effective retrieval. In this scenario, similarity learning approaches capable of improving the effectiveness of retrieval in an unsupervised way are indispensable. A novel method, called Cartesian Product of Ranking References (CPRR), is proposed with this objective in this paper. The proposed method uses Cartesian product operations based on rank information for exploiting the underlying structure of datasets. Only subsets of ranked lists are required, demanding low computational efforts. An extensive experimental evaluation was conducted considering various aspects, seven public multimedia datasets (images and videos) and several different features. Besides effectiveness, experiments were also conducted to assess the efficiency of the method, considering parallel and heterogeneous computing on CPU and GPU devices. The proposed method achieved significant effectiveness gains, including competitive state-of-the-art results on popular benchmarks. Keywords: Content-based image retrieval Unsupervised learning Cartesian product Effectiveness Efficiency (C) 2017 Elsevier B.V. All rights reserved. (AU)