Abstract: Binary segmentation for image collection has received considerable attention in image processing and graphic communities recently.Interactively separating foregrounds from an image set one by one is time-consuming and requires tedious user guidance.Meanwhile,image co-segmentation techniques generally lack efficiency due to the complexity of solving large optimization problems and are only applicable to images sharing similar foreground appearance.In this paper,we propose an example-driven semi-automatic framework to tackle the image collection segmentation problem.First,we select few sample images from the given image collection and deliver them to user for hand-segmentation.Then,super-pixel features based support vector machine(SVM) classifier is trained.For each super-pixel of a given image,we estimate its foreground labeling confidence by applying Sigmoid function on the distance between its descriptor and SVM separation hyperplane.The confidence values are then encoded in a graph cut segmentation procedure to achieve automatic object cutout.For each image with incorrectly segmented regions,accurate result is further obtained by a new proposed local refinement process.Experiments on 2 standard datasets are presented,showing that the proposed algorithm not only greatly outperforms existing co-segmentation techniques,but also largely reduces users' efforts for cutting out object interactively.