BEN: Using Confidence-Guided Matting for Dichotomous Image Segmentation

Current approaches to dichotomous image segmentation (DIS) treat imagematting and object segmentation as fundamentally different tasks. Asimprovements in image segmentation become increasingly challenging to achieve,combining image matting and grayscale segmentation techniques offers promisingnew directions for architectural innovation. Inspired by the possibility ofaligning these two model tasks, we propose a new architectural approach for DIScalled Confidence-Guided Matting (CGM). We created the first CGM model calledBackground Erase Network (BEN). BEN is comprised of two components: BEN Basefor initial segmentation and BEN Refiner for confidence refinement. Ourapproach achieves substantial improvements over current state-of-the-artmethods on the DIS5K validation dataset, demonstrating that matting-basedrefinement can significantly enhance segmentation quality. This work opens newpossibilities for cross-pollination between matting and segmentation techniquesin computer vision.