LVIS Challenge Track Technical Report 1st Place Solution: Distribution Balanced and Boundary Refinement for Large Vocabulary Instance Segmentation

This report introduces the technical details of the team FuXi-Fresher for\nLVIS Challenge 2021. Our method focuses on the problem in following two\naspects: the long-tail distribution and the segmentation quality of mask and\nboundary. Based on the advanced HTC instance segmentation algorithm, we connect\ntransformer backbone(Swin-L) through composite connections inspired by CBNetv2\nto enhance the baseline results. To alleviate the problem of long-tail\ndistribution, we design a Distribution Balanced method which includes dataset\nbalanced and loss function balaced modules. Further, we use a Mask and Boundary\nRefinement method composed with mask scoring and refine-mask algorithms to\nimprove the segmentation quality. In addition, we are pleasantly surprised to\nfind that early stopping combined with EMA method can achieve a great\nimprovement. Finally, by using multi-scale testing and increasing the upper\nlimit of the number of objects detected per image, we achieved more than 45.4%\nboundary AP on the val set of LVIS Challenge 2021. On the test data of LVIS\nChallenge 2021, we rank 1st and achieve 48.1% AP. Notably, our APr 47.5% is\nvery closed to the APf 48.0%.\n