BorderPointsMask: One-stage instance segmentation with boundary points representation

The mechanism of human vision can easily detect and segment objects based on boundary information. Even though the deep learning instance segmentation model based on boundary information can mimic this human vision mechanism, its prospect is rarely explored. In this work, we propose a one-stage and anchor-free instance segmentation framework based on boundary points representation, named BorderPointsMask. The proposed BorderPointsMask doesn't need to rely on the object detector. It formulates instance segmentation as instance classification, boundary points location prediction, and boundary points attribute prediction. Furthermore, we design two effective approaches to improve the performance of this proposed framework. Specifically, we utilize the proposed BorderPoints Center-ness to suppress the predicted low-quality masks. And the Deformation Before-and-After Stacking Module (DBASM) is formulated to promote instance classification and boundary points learning. BorderPointsMask obtains 35.0% in mask Average Precision (AP) with single-model (ResNet-101-FPN) and single-scale training/testing on the COCO benchmark, which demonstrates the superior performance among one-stage frameworks. The code will be available.