Cheng Tianheng ; Wang Xinggang ; Chen Shaoyu ; Zhang Qian ; Liu Wenyu

Abstract

Labeling objects with pixel-wise segmentation requires a huge amount of humanlabor compared to bounding boxes. Most existing methods for weakly supervisedinstance segmentation focus on designing heuristic losses with priors frombounding boxes. While, we find that box-supervised methods can produce somefine segmentation masks and we wonder whether the detectors could learn fromthese fine masks while ignoring low-quality masks. To answer this question, wepresent BoxTeacher, an efficient and end-to-end training framework forhigh-performance weakly supervised instance segmentation, which leverages asophisticated teacher to generate high-quality masks as pseudo labels.Considering the massive noisy masks hurt the training, we present a mask-awareconfidence score to estimate the quality of pseudo masks and propose thenoise-aware pixel loss and noise-reduced affinity loss to adaptively optimizethe student with pseudo masks. Extensive experiments can demonstrate theeffectiveness of the proposed BoxTeacher. Without bells and whistles,BoxTeacher remarkably achieves 35.0 mask AP and 36.5 mask AP with ResNet-50 andResNet-101 respectively on the challenging COCO dataset, which outperforms theprevious state-of-the-art methods by a significant margin and bridges the gapbetween box-supervised and mask-supervised methods. The code and models will beavailable at https://github.com/hustvl/BoxTeacher.