B$^{3}$CT: Three-branch Coordinated Training for Domain Adaptive Semantic Segmentation
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: domain adaptation, semantic segmentation
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Unsupervised domain adaptive semantic segmentation aims to adapt a dense prediction model trained on the source domain to the target domain by transferring knowledge without further annotations. A mainstream solution for transferring knowledge is to achieve alignment between different domains and eliminate domain gaps caused by source-target distributions. However, previous work paid little attention to where and when to align. We find that different contents in images are aligned at different stages of the whole network, and the alignment should be gradually strengthened during the whole training process due to the accuracy of target pseudo labels. Given these two observations, we propose a three-branch coordinated training (B$^{3}$CT) framework. Besides two normal source and target branches, a third branch is involved specifically for the alignment. In this branch, the hybrid-attention mechanism is utilized to do the alignment, while an Adaptive Alignment Controller (AAC) is built to adjust the contents being aligned according to the stages of the whole network. Meanwhile, in B$^{3}$CT, a coordinate weight is designed to gradually strengthen the importance of the alignment based on the training accuracy in the whole training process. Extensive experiments show that our proposed methods achieve competitive performances on tasks of GTA5$\to$Cityscapes and SYNTHIA$\to$Cityscapes.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3653
Loading