Feature Re-Representation and Reliable Pseudo Label Retraining for Cross-Domain Semantic Segmentation

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation is crucial for applying models to new datasets without manual labeling.
• Semantic segmentation models often suffer performance degradation when applied to domains different from their training data.

Purpose of the Study:

• To develop a novel unsupervised domain adaptation method for semantic segmentation.
• To effectively transfer knowledge from a source domain to a target domain while preserving target-specific information.

Main Methods:

• A novel feature re-representation strategy using source domain bases.
• Employing a discriminator to minimize the domain gap between source and target features.
• Introducing a Reliable Pseudo Label Retraining (RPLR) strategy for enhanced semantic meaningfulness.

Main Results:

• The proposed method successfully minimizes the domain gap, enabling effective knowledge transfer.
• Re-represented target features retain source domain knowledge and target-specific information.
• Extensive experiments show competitive performance on semantic segmentation benchmarks.

Conclusions:

• The novel unsupervised domain adaptation method significantly improves semantic segmentation performance.
• The approach effectively bridges the domain gap by re-representing features.
• The RPLR strategy enhances the semantic interpretability of adapted features.