Optimizing the decomposition for multiple foreground cosegmentation

摘要

The goal of multiple foreground cosegmentation (MFC) is to extract a finite number of foreground objects from an input image collection, while only an unknown subset of such objects is presented in each image. In this paper, we propose a novel unsupervised framework for decomposing MFC into three distinct yet mutually related tasks: image segmentation, segment matching, and figure/ground (F/G) assignment. By our decomposition, image segments sharing similar visual appearances will be identified as foreground objects (or their parts), and these segments will be also separated from background regions. To relate the decomposed outputs for discovering high-level object information, we construct foreground object hypotheses, which allows us to determine the foreground objects in each individual image without any user interaction, the use of pre-trained classifiers, or the prior knowledge of foreground object numbers. In our experiments, we first evaluate our proposed decomposition approach on the iCoseg dataset for single foreground cosegmentation. Empirical results on the FlickrMFC dataset will further verify the effectiveness of our approach for MFC problems.