Weakly Supervised Method for Domain Adaptation in Instance Segmentation

Jie Sun, Yan Tian, Jialei Wang, Zhaocheng Xu, Hao Wang, Zhaoyi Jiang, Xun Wang

Abstract: The domain adaptation of an instance segmentation model has gained much attention. However, manual annotation is tedious and self-training contains too much pseudolabel noise. Inspired by weakly supervised methods, we propose a method to handle these challenges by limited verification signals and label propagation. Semantic trees are constructed to explore the relation between samples by using a clustering method; Then, reliable pseudolabels are verified and propagated to unreliable labels, which improves instance segmentation model by employing the updated samples. Experiments on public datasets demonstrate that the proposed approach is competitive with state-of-the-art approaches.