Image Clustering Based on the Human Intelligence

Abstract

Current clustering algorithms mainly base on calculating distance between items that provides similarity information. But the distance cannot reflect all the correct information between items, which may lead to significant errors. We study the problem of seeking pairwise constraints with crowdsourcing in order to improve clustering results. Crowdsourcing is an emerging and powerful paradigm, which enables the use of background knowledge collecting from users, and image clustering is a relevant and appropriate use case. We propose a framework bringing in human intelligence during the clustering process. The key point of the framework is to choose best questions to perform on the crowdsourcing platform, gather pairwise constraints, and melt the existing algorithm and human input together. As the computation is extensive, we also provide some heuristic optimal methods, including natural transitive relations, to reduce the number of HITs of asking people. We evaluate the framework on real image dataset. The experiment result demonstrates the algorithm achieves a fairly good performance comparing to the other state-of-theart methods, and the optimized strategies significantly reduce the number of HIT.