Bidirectional Learning for Domain Adaptation of Semantic Segmentation
Lu Yuan2
UC San Diego1, Microsoft2
Overview
Domain adaptation for semantic image segmentation is very necessary since manually labeling large datasets with pixel-level labels is expensive and time consuming. Existing domain adaptation techniques either work on limited datasets, or yield not so good performance compared with supervised learning. In this project, we propose a novel bidirectional learning framework for domain adaptation of segmentation. Using the bidirectional learning, the image translation model and the segmentation adaptation model can be learned alternatively and promote to each other. Furthermore, we propose a self-supervised learning algorithm to learn a better segmentation adaptation model and in return improve the image translation model. Experiments show that our method is superior to the state-of-the-art methods in domain adaptation of segmentation with a big margin.
Published in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019.
Models
Architecture: Architecture of bidirectional learning network and loss functions.
Highlights
Forward direction: the images are translated to the target domain and forwarded to the segmentation model. Backward direction: the segmentation model is fixed and served as a perceptual loss for learning the image translation model.
Benefit
- The gap between source and target domain can be iteratively reduced.
- The framework can be transferred from source to target as well as from target to source.
Results
Qualitative Result: BDL transfers images from source to target as well as from target to source.
Qualitative Result: The segmentation results given by BDL.
