Co-Training for Unsupervised Domain Adaptation of Semantic Segmentation Models.

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Systems

Background:

• Semantic image segmentation is crucial for autonomous driving, but deep model training is hindered by extensive human-based image labeling.
• Synthetic data with automatic labels offers a viable alternative, but requires addressing unsupervised domain adaptation (UDA).

Purpose of the Study:

• To propose a new co-training procedure for unsupervised domain adaptation (UDA) in semantic segmentation, specifically for synth-to-real adaptation.
• To enhance the performance of deep models for autonomous driving by reducing the need for manually labeled real-world data.

Main Methods:

• A co-training procedure is introduced, involving iterative labeling of unlabeled real-world images by intermediate deep models.
• A self-training stage generates two domain-adapted models, followed by a collaboration loop for mutual improvement.
• The models treat deep networks as black boxes, collaborating at the pseudo-labeled target image level without modifying loss functions or explicit feature alignment.

Main Results:

• The proposed co-training procedure demonstrated significant improvements in semantic segmentation accuracy.
• Performance gains ranged from approximately 13 to 31 mean Intersection over Union (mIoU) points compared to baseline methods.
• The method was tested on standard synthetic and real-world datasets for onboard semantic segmentation.

Conclusions:

• The developed co-training strategy effectively bridges the domain gap between synthetic and real-world data for semantic segmentation.
• This approach offers a promising solution for training robust autonomous driving models with reduced manual annotation effort.
• The black-box model collaboration at the pseudo-labeling stage proves effective for UDA without complex modifications.