The Last Mile to Supervised Performance: Semi-Supervised Domain Adaptation for Semantic Segmentation
Abstract: Supervised deep learning requires massive labeled datasets, but obtaining annotations is not always easy or possible, especially for dense tasks like semantic segmentation. To overcome this issue, numerous works explore Unsupervised Domain Adaptation (UDA), which uses a labeled dataset from another domain (source), or Semi-Supervised Learning (SSL), which trains on a partially labeled set. Despite the success of UDA and SSL, reaching supervised performance at a low annotation cost remains a notoriously elusive goal. To address this, we study the promising setting of Semi-Supervised Domain Adaptation (SSDA). We propose a simple SSDA framework that combines consistency regularization, pixel contrastive learning, and self-training to effectively utilize a few target-domain labels. Our method outperforms prior art in the popular GTA$\rightarrow$Cityscapes benchmark and shows that as little as $50$ target labels can suffice to achieve near-supervised performance. Additional results on Synthia$\rightarrow$Cityscapes, GTA$\rightarrow$BDD and Synthia$\rightarrow$BDD further demonstrate the effectiveness and practical utility of the method. Lastly, we find that existing UDA and SSL methods are not well-suited for the SSDA setting and discuss design patterns to adapt them.
Submission Length: Regular submission (no more than 12 pages of main content)
Changes Since Last Submission: All changes are highlighted in red to facilitate in the manuscript, to make it easier to find them.
Particularly, following the first review, we have:
- Extended related work (Section 2) to provide more context on the current state of SSDA for segmentation and related fields (e.g., how does SSDA compare to UDA and SSL).
- Added a discussion on the comparison to previous works and the need for explicit domain mixing. (Section 4.2.1).
For the second review, we have
- Commented on the quality of psuedo-labels and clarified the use of a confidence threshold (Section 3.4).
- Commented on the convergence of the self-training algorithm (Section 4.3.2).
- Added a reference (Liu et al, 2022a) to related work.
For the third review we have
- Extended on hyperparameter tuning in Appendix B
Assigned Action Editor: ~Yale_Song1
Submission Number: 1617
Loading