Contour and Enclosed Region Refining for Contour-based Instance Segmentation

Current contour-based instance segmentation methods predict sets of vertexes to form contours to enclose object instances in images for realizing instance segmentation. Due to the inaccuracy of contour vertexes for describing object instances, mask decay and contour decay issues arise, limiting the performances of contour-based instance segmentation methods. In order to address these issues, in this paper we propose to design a contour and enclosed region refining network module, named CORE, to integrate to basic contour-based instance segmentation methods to obtain high-quality instance segmentation results. Specifically, we adopt a graph convolutional network to utilize correlation among initially predicted contour vertexes for refinement to address the contour decay issue. And, we predict and assemble a set of boundary-aware heatmaps to eliminate external regions enclosed within predicted object instance contours to relieve the mask decay problem. Furthermore, we propose several improvements that can be made to a basic contour-based instance segmentation method, i.e. Polar GIoU loss, Internal Center, and Hard Sample Polar Centerness. Finally, extensive experiments are conducted on the COCO dataset to evaluate the effectiveness of the proposed method. Experimental results show that our method can achieve 39.8 mAP on the COCO dataset, which outperforms state-of-the-art contour-based instance segmentation methods.