Real-time instance segmentation with assembly parallel task

Although instance segmentation has made significant progress in recent years, it is still a challenge to develop highly accurate algorithms with real-time performance. In this paper, we propose a real-time framework denoted by APTMask for instance segmentation, which builds on the real-time project YOLACT. In APTMask, we use Swin-Transformer Tiny with PA-FPN as the default feature backbone and a base image size of 544 x 544. We devise a new mask branch, which can more effectively exploit the semantic information of PA-FPN deeper features and the positional information of shallow features for mask representation, compared to the use of implicit parameterized forms. We replace fast NMS with Cluster NMS, which compensates for the performance penalty of fast NMS compiled to standard NMS. CIoU loss is also adopted to fully exploit the scale information of the aspect ratio of the bounding box. Experimental results show that APTMask can achieve 39.7/34.7 box/mask AP on COCO val2017 dataset at 31.8 fps evaluated with a single RTX 2080TI GPU card. Compared to YOLACT, APTMask improves the box AP by about 8.0% and the mask AP by 6.2%, which is encouraging and competitive. Given its simplicity and efficiency, we hope that our APTMask can serve as a simple but strong baseline for a variety of instance-wise prediction tasks.