LV Team Atop ICCV 2019 Large Vocabulary Instance Segmentation Challenge
Our lab members, including Tao Wang, Yu Li, Bingyi Kang, Jun Hao Liew, Sheng Tang, and Jiashi Feng, won the 1st Place Entry award at the 1st ICCV 2019 COCO Large Vocabulary Instance Segmentation (LVIS) Challenge at 10/27/2019 at Seoul, Korea.
A classification calibration method for long-tail instance segmentation was proposed to improve the inaccurate proposal classification for tail classes. The authors calibrate the prediction of classification head with a specifically designed twin head which is trained with proposal-level class balance sampling strategy. Without much additional cost and modification, this method improves the performance of baseline by a large margin on tail classes. Our team is one of the co-winners of the challenge, and we surpass the runner-up team by 2.7% in mAP.
Corresonding tech report has been posted on https://arxiv.org/abs/1910.13081, with an improved version submitted to CVPR 2020. See https://www.lvisdataset.org/challenge for more details.