Interactive segmentation based on super-pixel and multi-cues combination

Interactive segmentation is very useful in many computer vision applications, and in which graph cut is a very popular technique. Traditional graph cut approaches usually assign labels to pixels or pixel-grids. To large scale images, those approaches are very time consuming, and possibly fail when there are some similar properties between foreground and background in color, texture, etc. In this work, we improve the computation process of the edge costs in graph with neighborhood interactions, and propose a new edge descriptor to measure edge continuity among neighboring nodes. Rather than a simple combination of multi-cues, we use parameter learning to predict the weights of different cues. The experiment results show our method increased segmentation accuracy and reduced effort on the part of the user.