Image cosegmentation via multi-task learning

Image segmentation has been studied in computer vision for many years and yet it remains a challenging task. One major difficulty arises from the diversity of the foreground, which often results in ambiguity of background-foreground separation, especially when prior knowledge is missing. To overcome this difficulty, cosegmentation methods were proposed, where a set of images sharing some common foreground objects are segmented simultaneously. Different models have been employed for exploring such a prior of common foreground. In this paper, we propose to formulate the image cosegmentaion problem using a multi-task learning framework, where segmentation of each image is viewed as one task and the prior of shared foreground is modeled via the intrinsic relatedness among the tasks. Compared with other existing methods, the proposed approach is able to simultaneously segment more than two images with relatively low computational cost. The proposed formulation, with three different embodiments, is evaluated on two benchmark datasets, the CMU iCoseg dataset and the MSRC dataset, with comparison to leading existing methods. Experimental results demonstrate the effectiveness of the proposed method.