MADAv2: Advanced Multi-Anchor Based Active Domain Adaptation Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised domain adaptation is crucial for tasks with limited annotated data.
• Standard methods can distort target-domain data structure, reducing performance.
• Semantic segmentation often faces challenges due to domain shift.

Purpose of the Study:

• To improve unsupervised domain adaptation for semantic segmentation.
• To mitigate performance degradation caused by unconditional distribution mapping.
• To introduce a more effective sample selection strategy.

Main Methods:

• Proposed active sample selection using multiple anchors to characterize multimodal distributions.
• Developed a semi-supervised domain adaptation strategy to address long-tail distributions.
• Utilized innovative techniques to better represent source and target domains.

Main Results:

• Achieved significant performance gains by alleviating target-domain distribution distortion.
• Outperformed state-of-the-art methods on public datasets (GTA5, SYNTHIA).
• Reached performance comparable to fully-supervised methods (71.4% mIoU on GTA5, 71.8% mIoU on SYNTHIA).

Conclusions:

• Active sample selection effectively enhances unsupervised domain adaptation.
• The proposed semi-supervised strategy further boosts segmentation performance.
• The approach demonstrates superior effectiveness and robustness in semantic segmentation tasks.