Novel-domain Object Segmentation via Reliability-aware Teacher Ensemble

Object segmentation in a novel domain is challenging due to the scarcity of high-quality label data. A common practice is to leverage pseudolabels generated from a teacher model pretrained on benchmark datasets. However, such a teacher model may produce unreliable pseudolabels due to semantic shift from a standard domain to a novel domain. Blindly trusting pseudolabels may actually harm the target model's performance. To address this challenge, we propose a method to optimally adapt pseudolabels to work on a novel domain. At the center of our approach is a reliability score for pseudolabels that allows for an optimal mix of true labels and pseudolabels in training. Mirroring the standard practise of creating a benchmark dataset (such as ImageNet or GoEmotions), we use an ensemble of teacher models, instead of a single teacher model. Our reliability score captures both self-confidence of each teacher model and inter-model consistency among teacher models. This reliability score provides a disciplined way to balance the additional information in pseudolabels and the risk of overfitting incorrect labels. To test our approach for novel-domain object segmentation, we created an oil-leak surveillance dataset, which we have made publicly available. The experimental results show that our method achieves higher segmentation IoU over some strong baselines.