Intelligent Understanding of User Input Applied to Arc-Weight Estimation for Graph-Based Foreground Segmentation

We present an intelligent approach for understanding user interaction to simplify the interface of graph-based image segmentation. The user draws a set of markers (strokes) over the object and the background, and our method automatically determines a subset of these pixels with dissimilar image properties for arc-weight estimation. Arc-weight estimation combines object information learned from the set of selected pixels with image information, to make object delineation more effective. Our method differs from approaches that recompute the arc weights carelessly during delineation, by further interpreting user interaction to determine when and where to recompute them. Furthermore, we build our framework around the image foresting transform (IFT), by taking advantage of its operators for supervised fuzzy classification, clustering, and object delineation. We evaluate our framework using a dataset with 50 natural images and by comparing it against another recent IFT-based method, which computes arc weights in a separated step of user interaction for more effective segmentation.