Image Co-Segmentation via Locally Biased Discriminative Clustering

Abstract

Object co-segmentation aims at simultaneously extracting common objects appeared in multiple images. In this paper, we propose a novel object co-segmentation method in which we formulate the image co-segmentation as a locally biased discriminative clustering problem. Specifically, we add a seed vector and a constraint term into the framework of discriminative clustering to constrain the segmentation result bias to this seed vector. In order to deal with the co-segmentation problem with indefinite number of common foreground objects, we design a Markov Random Field (MRF) based method to extract common objects prior. The extracted common objects prior is then added into the discriminative clustering and treated as the seed vector to constrain the segmentation result biased to it. Under the supervision of this prior, the segmentation result can find more common objects and becomes more complete and meaningful. In addition, we present a new segmentation process where we alternately update the MRF-based model and discriminative clustering to refine the common objects prior and the segmentation result. Finally, we test the proposed method on three benchmark datasets, iCoseg, Coseg-Rep, and THUR15K. The experimental results demonstrate that the proposed method outperforms other state-of-the-art methods.