A modular approach to image matting

Abstract

Image matting is the art of creating an accurate alpha matte for the purpose of foreground separation in an image or video. Although there have been many methods which only require an input image, the best-performing image matting models continue to rely on additional inputs — mainly the trimap — for more accurate alpha matte estimations. We propose a modular image matting architecture which leverages advancements in semantic segmentation and our trimap generation network to allow for a trimap-free approach to some of the most popular trimap-based image matting methods. Our design delivers promising results, allowing users to take advantage of powerful trimap-based methods, without having to worry about additional inputs, all while granting them the freedom to swap different networks in and out for the different stages of the modular architecture.