Multi-class Cosegmentation with Pairwise Active Learning

Abstract

Jointly segmenting common objects from multiple images remains a challenging problem. In this paper, we propose a multi-class cosegmentation method based on correlation clustering, which requires no prior knowledge of the number of clusters. Our method can handle large number of images because of the flexible graph structure and scalable clustering method. Moreover, we use active learning to intelligently recommend pairs of regions to users, in order to get pairwise must-link and cannot-link constraints. Then a novel dimensionality reduction method is proposed to produce an affinity matrix which reflects both the intrinsic structure of data and the constraints. Finally, correlation clustering is applied on the newly generated affinity matrix to acquire refined results. Experimental results show that our system can correct errors of initial segmentation and personalize segmentation result according to user preferences.