Image matting for sparse user input by iterative refinement

Image matting is the process of extracting theforeground component from an image. Since matting is an underconstrained problem most techniques address the case whereusers supply some dense labelling to indicate known foregroundand background regions. In contrast to other techniques ourproposed technique is unique in that focuses on achievingsatisfactory results with extremely sparse input, e.g. a handfulof individual pixel labels. We propose an iterative extension tothe class of affinity matting techniques. Analysis of results fromaffinity matting with sparse labels reveals that the low qualityalpha mattes can be processed and re-used for the next iteration.We demonstrate this extension using the recent KNN matting andshow that this technique can greatly improve matting results.