A Real-time Interactive Tool for Image Cutout

Abstract

We present an interactive tool for extracting foreground objects from image in real-time. As seen from the object boundary, the segmentation is to split pixel-pairs right on the boundary. The system utilizes the user input while the user roughly paints the stroke along the object boundary. Either side of the stroke is assumed to lie in foreground or background, and the mouse trace is supposed to coarsely indicate the object contour. We use the appearance, gradient and contour information to formulate the segmentation problem to an energy minimization problem, which is solved by graph cut. The appearance model is built via local sampling, since it often provides more discriminative model in local areas than that from global sampling. In addition, our system can dynamically adjust the stroke and the optimization region for which segmentation needs to be computed. The optimization region is much smaller than the whole image. This greatly reduces the computational complexity in each iteration and also gives the system the ability of incrementally segmenting. Experiments show the effectiveness of our methods in improving the segmentation performance.