Research map / Label-efficient and robust learning / Domain adaptation and generalization
Domain-Agnostic Mutual Prompting for Unsupervised 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.
Proposes DAMP, which freezes CLIP and mutually prompts its text and vision branches via cross-attention to learn domain-agnostic, instance-conditioned embeddings for unsupervised domain adaptation.
Problem
Unsupervised domain adaptation with vision-language models: existing CLIP prompt-based UDA methods learn domain-specific textual prompts and prompt only one modality, limiting cross-domain knowledge transfer.
Approach
Mutual prompting between frozen CLIP text and visual branches through a cross-attention decoder: image context adapts class text prompts per instance, while the domain-agnostic text prompts guide visual prompts; regularized by a semantic-consistency loss, an instance-discrimination contrastive loss, and information maximization.
Key results
- On Office-Home with ResNet-50, DAMP reaches 78.2% average accuracy, versus 74.5% for DAPrompt and 76.6% for PADCLIP's ViT-free baseline row (PADCLIP 76.6%).
- On Mini-DomainNet with ResNet-50, DAMP achieves 77.5% average, improving over CLIP (71.2%) by 6.3 points and AD-CLIP (75.2%) by 2.3 points.
- On VisDA-17 with ResNet-101, DAMP gets 88.4% versus PADCLIP's 88.5%, while tuning only about 11.9% of PADCLIP's parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DAMP instead of PADCLIP | ResNet-50 · Office-Home | accuracy: 76.6 -> 78.2 | Helped |
| Use DAMP instead of PADCLIP | ViT-B/16 · Office-Home | accuracy: 86.7 -> 87.1 | Helped |
| Use DAMP instead of PADCLIP | ResNet-101 · VisDA-17 | accuracy: 88.5 -> 88.4 | Hurt |
| Use DAMP instead of PADCLIP | ViT-B/16 · VisDA-17 | accuracy: 90.9 -> 90.9 | No clear effect |
| Use DAMP instead of AD-CLIP | ResNet-50 · Mini-DomainNet | accuracy: 75.2 -> 77.5 | Helped |
| Use DAMP instead of AD-CLIP | ViT-B/16 · Mini-DomainNet | accuracy: 86.9 -> 87.6 | Helped |
| Add visual prompting only (ablation) | CoOp baseline · Office-Home | accuracy: 75.3 -> 75.7 | Helped |
| Add visual prompting only (ablation) | CoOp baseline · VisDA-17 | accuracy: 86.1 -> 86.5 | Helped |
| Add instance-level textual prompting only (ablation) | CoOp baseline · Office-Home | accuracy: 75.3 -> 75.8 | Helped |
| Add instance-level textual prompting only (ablation) | CoOp baseline · VisDA-17 | accuracy: 86.1 -> 87.1 | Helped |
| Add mutual instance-level textual and visual prompting (ablation) | CoOp baseline · Office-Home | accuracy: 75.3 -> 76.1 | Helped |
| Add mutual instance-level textual and visual prompting (ablation) | CoOp baseline · VisDA-17 | accuracy: 86.1 -> 87.2 | Helped |
Surprising findings
- Disabling the source supervision loss in DAPrompt has only marginal influence on target performance, suggesting source knowledge is largely encoded in source-specific prompts that do not transfer.
Limitations
- Results on VisDA-17 with ResNet-101 are slightly below PADCLIP (88.4% vs 88.5%)
- Performance sensitive to the learnable-token length and loss-weight choices in some settings, per the hyperparameter analysis
Experimental setup
Datasets: Office-Home VisDA-17 Mini-DomainNet
Benchmarks: Office-Home VisDA-17 Mini-DomainNet
Metrics: classification accuracy (%)
Baselines: DAPrompt AD-CLIP PADCLIP CLIP SSRT TVT CDTrans SRDC ToAlign CoOp
Builds on: CoOp CoCoOp MaPLe DAPrompt FixMatch
Closest papers by meaning
- Prompt-Based Distribution Alignment for Unsupervised Domain Adaptation (AAAI 2024 · similarity 0.92)
- Enhancing Domain Adaptation through Prompt Gradient Alignment (NeurIPS 2024 · similarity 0.92)
- Preserving Clusters in Prompt Learning for Unsupervised Domain Adaptation (CVPR 2025 · similarity 0.92)
- Domain-Controlled Prompt Learning (AAAI 2024 · similarity 0.90)
- Exploring Sparse Visual Prompt for Domain Adaptive Dense Prediction (AAAI 2024 · similarity 0.90)
- PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024 · similarity 0.90)
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.