Abstract
Instance segmentation is a crucial task in computer vision that aims to simultaneously identify and segment individual objects within images. While existing approaches such as Mask R-CNN have shown promise, they often struggle with accurate boundary detection, especially for complex objects. In this paper, we introduce BorderMask, a novel framework that enhances boundary perception and streamlines instance segmentation. BorderMask comprises three key innovations: the Multiscale Boundary Perception Enhanced Attention (MBPEA) module, which iteratively optimizes multi-scale boundary features; the Cross-modal Link Structure (CMLS), which enables information exchange between detection and segmentation branches; and the Equilibrium Map loss function, which mitigates class imbalance issues. Extensive experiments on benchmark datasets including MS COCO, PASCAL VOC 2012, and Cityscapes demonstrate that BorderMask significantly outperforms state-of-the-art methods, achieving an AP of 44.7% on MS COCO, underscoring its robustness and effectiveness. The code will be available at https://gitee.com/shi-junyong/BorderMask .