High Quality Mask Generation Using HardRegion Refinement in Object-based Instance Segmentation

Existing object-based instance segmentation methods have made remarkable achievements recently. However, the conventional methods ignorethe inhomogeneity of the segmentation difficulty in the spatial dimension in different regions of instances. In this paper, we propose theHard Region Refinement (HRR) method to explicitly model easy andhard regions adaptively for high-quality instance masks generating.Our insight is that different regions of an instance should be treatedunequally on account of their different complexity. This approach decomposes the instance prediction into coarse prototype mask estimation andmultiple fine-grained hard mask prediction. To supervise the regionsat high resolution, Region-DCT (R-DCT) loss is proposed for theefficient region learning. We have carried out extensive exploratoryexperiments on the COCO and Cityscapes datasets. The experimental results demonstrate the superiority of our proposed method.