Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Unified Language-driven Zero-shot Domain Adaptation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Unified Language-driven Zero-shot Domain Adaptation (ULDA), where one segmentation model adapts to multiple unseen target domains from text descriptions alone, without domain IDs, via hierarchical context alignment, domain consistent representation learning, and a text-driven rectifier.
Problem
Existing language-driven zero-shot domain adaptation (PØDA) needs domain IDs to pick a domain-specific segmentation head, which limits flexibility and scalability when adapting to several target domains without target images.
Approach
Align simulated target features with CLIP text embeddings at scene, regional, and pixel levels; preserve inter-class correlations across domains with a consistency loss; and rectify PIN-simulated features with text-derived statistics during head fine-tuning, so one shared head serves all domains.
Key results
- On Cityscapes to ACDC clear-to-adverse weather, ULDA reaches 42.47% mean mIoU, versus 40.65% for PØDA with domain-specific heads.
- On synthetic-to-real (GTA5 to Cityscapes+ACDC), ULDA reaches 33.21% mean mIoU, versus 31.43% for PØDA*.
- Ablation on Cityscapes-to-ACDC shows HCA, DCRL, TDR and segmentation loss together raise mean mIoU from 39.67% to 42.47%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ULDA instead of PØDA | CLIP-ResNet-50 segmentation model · ACDC Night | mIoU: 25.03 -> 25.40 | Helped |
| Use ULDA instead of PØDA | CLIP-ResNet-50 segmentation model · ACDC Snow | mIoU: 43.90 -> 46.00 | Helped |
| Use ULDA instead of PØDA | CLIP-ResNet-50 segmentation model · ACDC Rain | mIoU: 42.31 -> 44.94 | Helped |
| Use ULDA instead of PØDA | CLIP-ResNet-50 segmentation model · GTA5 | mIoU: 40.77 -> 42.91 | Helped |
| Use ULDA instead of PØDA | CLIP-ResNet-50 segmentation model · Cityscapes | mIoU: 40.02 -> 41.73 | Helped |
| Use ULDA with a shared head instead of PØDA* with separate heads | Cityscapes source model on ACDC adverse-weather domains · ACDC adverse weather | Mean-mIoU: 40.65 -> 42.47 | Helped |
| Use ULDA with a shared head instead of PØDA* with separate heads | GTA5 source model on Cityscapes and ACDC · Cityscapes and ACDC | Mean-mIoU: 31.43 -> 33.21 | Helped |
| Add HCA (ablation) | Baseline language-driven adaptation on Cityscapes→ACDC · ACDC adverse weather | Mean-mIoU: 39.67 -> 40.52 | Helped |
| Add DCRL to HCA (ablation) | HCA adaptation on Cityscapes→ACDC · ACDC adverse weather | Mean-mIoU: 40.52 -> 41.49 | Helped |
| Add TDR to HCA (ablation) | HCA adaptation on Cityscapes→ACDC · ACDC adverse weather | Mean-mIoU: 40.52 -> 41.35 | Helped |
| Add TDR to HCA + DCRL (ablation) | HCA + DCRL adaptation on Cityscapes→ACDC · ACDC adverse weather | Mean-mIoU: 41.49 -> 41.68 | Helped |
| Add segmentation loss (ablation) | HCA + DCRL + TDR adaptation on Cityscapes→ACDC · ACDC adverse weather | Mean-mIoU: 41.68 -> 42.47 | Helped |
Surprising findings
- A single shared head trained with ULDA outperforms PØDA's four domain-specific heads on the clear-to-adverse benchmark.
Limitations
- Relies on simple one-sentence prompts per domain and a frozen-backbone head-tuning setup; generality beyond driving scenes is untested in the main text.
- TDR cannot be applied in Stage-1 because of a trivial solution, so the rectifier is limited to head fine-tuning.
Experimental setup
Datasets: Cityscapes ACDC GTA5
Benchmarks: ACDC Cityscapes GTA5
Metrics: mIoU mAcc
Baselines: CLIPStyler PØDA source-only
Builds on: PØDA CLIP CLIPstyler DeepLabv3+
Closest papers by meaning
- Target Semantics Clustering via Text Representations for Robust Universal Domain Adaptation (AAAI 2025 · similarity 0.90)
- Source-Free Domain Adaptation with Frozen Multimodal Foundation Model (CVPR 2024 · similarity 0.88)
- Enhancing Domain Adaptation through Prompt Gradient Alignment (NeurIPS 2024 · similarity 0.87)
- Domain-Agnostic Mutual Prompting for Unsupervised Domain Adaptation (CVPR 2024 · similarity 0.87)
- Prompt-Based Distribution Alignment for Unsupervised Domain Adaptation (AAAI 2024 · similarity 0.87)
- Open-Set Heterogeneous Domain Adaptation: Theoretical Analysis and Algorithm (AAAI 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.