Uncertainty-Aware Source-Free Domain Adaptive Semantic Segmentation.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Source-Free Domain Adaptation (SFDA) addresses distribution shifts between training and deployment data without requiring source data.
• Semantic segmentation often relies on pseudo-labeling for target domain self-training, but source model-generated pseudo-labels can be unreliable due to domain shift.

Purpose of the Study:

• To enhance Source-Free Domain Adaptation for semantic segmentation by improving the reliability of pseudo-labels.
• To introduce a novel approach using Bayesian Neural Networks (BNN) to estimate and exploit pseudo-label uncertainty.

Main Methods:

• Utilized Bayesian Neural Networks (BNN) to quantify uncertainty in pseudo-labels generated for the target domain.
• Developed two novel self-training components: Uncertainty-aware Online Teacher-Student Learning (UOTSL) and Uncertainty-aware FeatureMix (UFM).
• Implemented and evaluated the proposed methods on GTA 5 → Cityscapes and SYNTHIA → Cityscapes benchmarks.

Main Results:

• The proposed BNN-based approach significantly improved the quality of pseudo-labels by effectively estimating their uncertainty.
• Achieved substantial performance gains, with mIoU increases of 3.6% on GTA 5 → Cityscapes and 5.7% on SYNTHIA → Cityscapes.
• Demonstrated superiority over existing state-of-the-art methods in Source-Free Domain Adaptation for semantic segmentation.

Conclusions:

• Bayesian Neural Networks offer a robust mechanism for uncertainty estimation in SFDA, leading to more reliable pseudo-labels.
• The proposed UOTSL and UFM components effectively leverage uncertainty information to boost semantic segmentation performance.
• This work presents a significant advancement in SFDA for semantic segmentation, particularly in scenarios with limited or no source data.