Improved PolarMask with Attention for Instance Segmentation

Abstract

As an important field of computer vision, instance segmentation is mainly divided into one-stage segmentation methods and two-stage segmentation methods. PolarMask is a one-stage instance segmentation method, which has the advantages of simple structure and fast speed. However, its accuracy is not good enough. To optimize the accuracy of PolarMask, we firstly propose attention-based polar Intersection over Union (IoU) loss based on PolarMask, and then we replace Feature Pyramid Network (FPN) structure and IoU loss with FPN-involution, PAFPN, and CIoU loss, respectively. Through ablation study, it is proved the effect of the proposed attention-based Polar IoU loss and verified the effect of the replacement module in the model. Further, through combination experiments, we find the most efficient combination method on COCO minival dataset. Finally, we compare our method with others, and an Improved PolarMask method is obtained.