Dichotomous Image Segmentation

Dichotomous image segmentation (DIS) focuses on achieving highly accurate, pixel-level segmentation of foreground objects from complex backgrounds in high-resolution images, a task challenging due to fine details and varied object appearances. Current research emphasizes improving model architectures, such as encoder-decoder networks with multi-view aggregation or bilateral reference mechanisms, to better capture both global context and local features, often incorporating auxiliary supervision strategies. This work is significant for advancing computer vision capabilities in applications like background removal, 3D reconstruction, and medical image analysis, where precise object delineation is crucial. The development of large-scale datasets and new evaluation metrics, such as human correction effort, further enhances the rigor and impact of DIS research.