Pseudolabel guided pixels contrast for domain adaptive semantic segmentation

2024-12-30

Scientific Reports

AI Summary

To address the bias in pixel-wise contrastive learning caused by noisy pseudo-labels in unsupervised domain adaptation (UDA) for semantic segmentation, this paper proposes a dynamically confidence-weighted cross-domain pixel contrastive learning framework. The core innovation lies in the first integration of pseudo-label confidence modeling into pixel-level contrastive learning, enabling a confidence-aware positive/negative sample construction strategy that enhances both discriminability and consistency of cross-domain features. The method comprises four modules: pseudo-label generation, confidence estimation, confidence-weighted contrastive learning, and domain alignment. Evaluated on the standard GTAV→Cityscapes UDA benchmark, it achieves 72.3% mIoU—surpassing the state-of-the-art by 1.8%—and significantly mitigates contrastive learning mismatch induced by unreliable pseudo-labels.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Computer Vision: Adversarial Attacks & Robustness
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Problem

• Unsupervised Domain Adaptation
• Contrastive Learning
• Object Recognition

Innovation

• Pseudo Label Guided Pixel Contrast (PGPC)
• Unsupervised Domain Adaptation (UDA)
• Semantic Segmentation Enhancement