Interactive object segmentation using kernel density estimation based graph cuts

This paper proposes an efficient and flexible interactive object segmentation approach using kernel density estimation based graph cuts. First, the user draws scribbles to roughly mark the interested object and background, respectively, and the likelihood of object versus background is evaluated for each pixel using nonparametric kernel density estimation. Then pixels are globally classified into object and background using graph cuts, which uses likelihood as data cost and gradient information to generate a spatial varying smoothness cost. If the user is not satisfied with the initially segmented object, the user is allowed to mark these undesirable regions by drawing additional scribbles. A local region for refinement is then adaptively determined, and pixel reclassification is performed to extract a more accurate object boundary. Experimental results on a variety of images demonstrate that interested objects with good visual quality can be extracted with less user interaction and a timely response.