Vision-Language-Guided Pseudo-Labels for Unsupervised Domain Adaptation in Semantic Segmentation for Waste Sorting
A cross-modal pseudo-labeling pipeline using SAM for region proposals and EVA-CLIP for semantic labeling enables unsupervised domain adaptation without target-domain annotations. Confidence filtering ensures only reliable pseudo-labels are used for self-training. BLIP optionally provides language-grounded verification for ambiguous regions. Evaluated on synthetic-to-real autonomous driving and lab-to-factory industrial waste sorting, the pipeline consistently improves over source-only baselines. Pseudo-label quality, not quantity, is decisive in self-training under domain shift.
Researchers present a vision-language-guided pseudo-labeling pipeline for unsupervised domain adaptation in semantic segmentation, targeting industrial waste sorting and autonomous driving. The method combines SAM-generated region proposals with EVA-CLIP semantic labels, filtered by confidence, and optionally verified by BLIP. Experiments show consistent improvement over source-only baselines, emphasizing that pseudo-label quality is more important than quantity.
The pipeline leverages cross-modal alignment between visual regions and text descriptions to generate high-quality pseudo-labels without target annotations. Confidence filtering and optional BLIP verification reduce label noise, which is critical for self-training under domain shift. This approach demonstrates that foundation models can be combined to automate annotation in deployment-critical scenarios.
Industrial waste sorting and autonomous driving face high labeling costs and domain gaps. This method offers a practical path to reliable automatic annotation, potentially reducing the need for expensive manual labeling and accelerating deployment of segmentation models in real-world settings.
Reduces annotation costs and time for deploying semantic segmentation in industrial applications. Enables faster adaptation to new environments without labeled data, improving scalability and ROI for AI-driven quality control and autonomous systems.
Next signals include application to other domain shifts, integration with additional foundation models, and evaluation on larger-scale industrial datasets. The emphasis on pseudo-label quality may influence future research in semi-supervised and unsupervised domain adaptation.