Research map / Label-efficient and robust learning / Domain adaptation and generalization
Semantic Feature Learning for Universal Unsupervised Cross-Domain Retrieval
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 Universal Unsupervised Cross-Domain Retrieval (U²CDR), where query and retrieval domains may have different label spaces, and proposes the UEM two-stage framework (prototype-based contrastive intra-domain learning, semantic-preserving adversarial alignment, and switchable nearest-neighbor matching).
Problem
Unsupervised cross-domain retrieval assumes identical category spaces across query and retrieval domains, which is unrealistic; the paper targets retrieval when label spaces differ (closed, partial, and open-set settings) without any labels.
Approach
Builds a cross-domain unified prototypical structure via instance-prototype-mixed contrastive loss plus a semantic-enhanced loss, then performs domain alignment with a pairwise-correlation-preserving regularizer so prototypes stay stable, and retrieves neighbors using prototype-aware switchable matching.
Key results
- Closet setting: 81.9 average mAP@All on Office-31 vs 79.3 for the best baseline (DG-UCDIR); 52.4 vs 49.3 on Office-Home.
- Partial setting: 63.0 average mAP@All on Office-31 vs 51.0 for the best baseline (CoDA), a margin of about 12 points.
- Open-set DomainNet: 34.8 shared-set mAP@All and 80.9 open-set accuracy on average, vs 29.3 and 64.1 for the best baselines (UCDIR).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add prototype merging across domains (ablation) | UEM · DomainNet open-set UCDR | Shared-set mAP@All: 30.9 -> 34.8 | Helped |
| add SEL (ablation) | UEM · DomainNet open-set UCDR | Shared-set mAP@All: 33.0 -> 34.8 | Helped |
| use semantic-preserving domain alignment instead of standard domain adversarial learning (ablation) | UEM · DomainNet open-set UCDR | Shared-set mAP@All: 24.4 -> 34.8 | Helped |
| use SN²M instead of UCDIR matching (ablation) | UEM · DomainNet open-set UCDR | Shared-set mAP@All: 33.3 -> 34.8 | Helped |
Limitations
- Source code and implementation details are promised 'soon' but not released in the paper.
- Open-set detection procedure is deferred to unspecified future materials.
- DomainNet uses only 7 classes selected by a sample-count threshold, limiting evaluation scope.
Experimental setup
Datasets: Office-31 Office-Home DomainNet
Benchmarks: Office-31 Office-Home DomainNet
Metrics: mAP@All open-set detection accuracy
Baselines: CDS PCS DARLING DN2A UCDIR CoDA DG-UCDIR
Builds on: MoCo MoCo v2 domain adversarial learning prototypical contrastive learning UCDIR
Closest papers by meaning
- Unsupervised Cross-Domain Image Retrieval via Prototypical Optimal Transport (AAAI 2024 · similarity 0.89)
- Retrieval Across Any Domains via Large-scale Pre-trained Model (ICML 2024 · similarity 0.88)
- RoDA: Robust Domain Alignment for Cross-Domain Retrieval Against Label Noise (AAAI 2025 · similarity 0.88)
- Target Semantics Clustering via Text Representations for Robust Universal Domain Adaptation (AAAI 2025 · similarity 0.87)
- Effective Comparative Prototype Hashing for Unsupervised Domain Adaptation (AAAI 2024 · similarity 0.86)
- U-MARVEL: Unveiling Key Factors for Universal Multimodal Retrieval via Embedding Learning with MLLMs (ICLR 2026 · similarity 0.86)
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.