Research map / Label-efficient and robust learning / Domain adaptation and generalization
Unsupervised Cross-Domain Image Retrieval via Prototypical Optimal Transport
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 ProtoOT, an optimal transport formulation with K-means-derived class marginals and contrastive losses that jointly handles intra-domain representation learning and cross-domain alignment for unsupervised cross-domain image retrieval.
Problem
Unsupervised cross-domain image retrieval (UCIR) must retrieve same-category images across domains without labels, under severe intra- and cross-domain distribution imbalance, which standard and unbalanced OT handle poorly.
Approach
Replaces the uniform marginal in standard OT with class marginals approximated by K-means cluster proportions and uses K-means centroids as initial prototypes, then applies one ProtoOT formulation to both intra-domain clustering and cross-domain alignment, with contrastive losses that enforce local consistency and global discrimination.
Key results
- On DomainNet, ProtoOT averages P@200 of 57.26% versus 39.09% for the best baseline DD, a +18.17 point gain.
- On Office-Home, ProtoOT averages P@15 of 37.98%, +3.83 over DD (34.15%).
- Ablation on DomainNet: replacing both components with standard OT yields 46.58 P@200 vs 57.26 for full ProtoOT in the averaged 12-task table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ProtoOT instead of DD | Unsupervised cross-domain retrieval · DomainNet (average) | P@200: 39.09 -> 57.26; +18.17 | Helped |
| Use ProtoOT instead of DD | Unsupervised cross-domain retrieval · Office-Home (average) | P@15: 34.15 -> 37.98; +3.83 | Helped |
| Add cross-domain alignment (ablation) | ProtoOT intra-domain representation learning · DomainNet (average, 12 tasks) | P@200: 47.02 -> 57.26 | Helped |
| Use ProtoOT rather than standard OT for intra-domain learning (ablation) | Standard OT cross-domain alignment · DomainNet (average, 12 tasks) | P@200: 52.16 -> 57.26 | Helped |
| Use ProtoOT rather than standard OT for cross-domain alignment (ablation) | ProtoOT intra-domain learning · DomainNet (average, 12 tasks) | P@200: 55.43 -> 57.26 | Helped |
| Use ProtoOT for both intra-domain learning and cross-domain alignment rather than standard OT (ablation) | Unsupervised cross-domain retrieval · DomainNet (average, 12 tasks) | P@200: 51.46 -> 57.26 | Helped |
| Use ProtoOT rather than Unbalanced OT for both stages (ablation) | Unsupervised cross-domain retrieval · DomainNet (average, 12 tasks) | P@200: 52.47 -> 58.26 | Helped |
| Use ProtoOT rather than Partial OT for both stages (ablation) | Unsupervised cross-domain retrieval · DomainNet (average, 12 tasks) | P@200: 53.23 -> 58.26 | Helped |
| Set λ to 0.01 rather than 0.001 (ablation) | ProtoOT · DomainNet (average, 12 tasks) | P@200: 54.49 -> 57.26 | Helped |
| Set λ to 0.01 rather than 0.005 (ablation) | ProtoOT · DomainNet (average, 12 tasks) | P@200: 54.74 -> 57.26 | Helped |
| Set λ to 0.01 rather than 0.05 (ablation) | ProtoOT · DomainNet (average, 12 tasks) | P@200: 54.81 -> 57.26 | Helped |
| Set λ to 0.01 rather than 0.1 (ablation) | ProtoOT · DomainNet (average, 12 tasks) | P@200: 54.28 -> 57.26 | Helped |
Limitations
- Experiments limited to two datasets and a small subset of 7 DomainNet categories.
- Relies on K-means to estimate class marginals, whose quality under severe imbalance is not analyzed in depth.
Experimental setup
Datasets: Office-Home DomainNet ImageNet
Benchmarks: Office-Home DomainNet
Metrics: precision at k (P@1/5/15 on Office-Home; P@50/100/200 on Dom
Baselines: ID ProtoNCE CDS PCS DD Unbalanced OT Partial OT
Builds on: SwAV MoCov2 ProtoNCE Sinkhorn-Knopp
Closest papers by meaning
- Semantic Feature Learning for Universal Unsupervised Cross-Domain Retrieval (NeurIPS 2024 · similarity 0.89)
- Instance-Level Composed Image Retrieval (NeurIPS 2025 · similarity 0.87)
- Retrieval Across Any Domains via Large-scale Pre-trained Model (ICML 2024 · similarity 0.86)
- RoDA: Robust Domain Alignment for Cross-Domain Retrieval Against Label Noise (AAAI 2025 · similarity 0.86)
- DiDA: Disambiguated Domain Alignment for Cross-Domain Retrieval with Partial Labels (AAAI 2024 · similarity 0.86)
- Effective Comparative Prototype Hashing for Unsupervised Domain Adaptation (AAAI 2024 · 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.