Research map / Label-efficient and robust learning / Domain adaptation and generalization
DiDA: Disambiguated Domain Alignment for Cross-Domain Retrieval with Partial Labels
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 DiDA, a method for cross-domain image retrieval under partial labels that disambiguates candidate labels while aligning domains via prototype-score unitization and prototype-based domain alignment losses.
Problem
Cross-domain image retrieval when training samples carry partial labels (candidate sets containing the true label), where ambiguous supervision and domain gaps jointly hurt retrieval performance.
Approach
Treats prototypes as class-certain supervision: prototypes from both domains are passed through the classifier to produce prototype-scores pushed toward a target unit matrix (PSUL), which simultaneously disambiguates labels and narrows the domain gap; an additional prototype-similarity discrepancy loss (PBDA) further aligns domains.
Key results
- On OfficeHome at partial rate 0.3, DiDA reaches 0.591 mAP (last epoch) versus 0.268-0.467 for PLL baselines such as PICO and DPLL.
- Average mAP drops only 0.118 from partial rate 0.1 to 0.3 on OfficeHome, while PICO drops 0.345, DPLL 0.203 and PaPi 0.225.
- On Office31 at partial rate 0.3, DiDA achieves 0.886 mAP (last epoch), exceeding PICO by about 0.073 (per the text).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiDA instead of PICO | Cross-domain retrieval with partial labels · OfficeHome, partial rate 0.1 | mAP: 0.673 -> 0.704 | Helped |
| Use DiDA instead of DPLL | Cross-domain retrieval with partial labels · OfficeHome, partial rate 0.3 | mAP: 0.467 -> 0.591 | Helped |
| Use DiDA instead of PICO | Cross-domain retrieval with partial labels · Office31, partial rate 0.3 | mAP: 0.809 -> 0.886 | Helped |
| Use DiDA instead of PICO | Cross-domain retrieval with partial labels · ImageCLEF, partial rate 0.4 | mAP: 0.727 -> 0.762 | Helped |
| Add Lpsu (prototype-score unitization) (ablation) | DiDA · OfficeHome, partial rate 0.3 | mAP: 0.477 -> 0.573 | Helped |
| Add Lpsu (prototype-score unitization) (ablation) | DiDA · Office31, partial rate 0.3 | mAP: 0.788 -> 0.868 | Helped |
| Add Lpda (prototype-based domain alignment) (ablation) | DiDA with Lcls and Lpsu · OfficeHome, partial rate 0.3 | mAP: 0.573 -> 0.591 | Helped |
| Add Lpda (prototype-based domain alignment) (ablation) | DiDA with Lcls and Lpsu · Office31, partial rate 0.3 | mAP: 0.868 -> 0.886 | Helped |
| Add Lpsu (prototype-score unitization) (ablation) | DiDA with Lcls and Lpda · OfficeHome, partial rate 0.3 | mAP: 0.503 -> 0.591 | Helped |
| Add Lpsu (prototype-score unitization) (ablation) | DiDA with Lcls and Lpda · Office31, partial rate 0.3 | mAP: 0.828 -> 0.886 | Helped |
| Vary α within [5, 10] (ablation) | DiDA · OfficeHome A-C and Office31 D-A | mAP | No clear effect |
| Vary β within [0.01, 0.1] (ablation) | DiDA · OfficeHome A-C and Office31 D-A | mAP | No clear effect |
Surprising findings
- Many partial-label learning methods underperform unsupervised cross-domain retrieval baselines at high partial rates, indicating domain discrepancy strongly hurts them.
- At low partial rates on some OfficeHome tasks, DiDA surpasses the fully supervised cross-entropy baseline (e.g., R-P task).
Limitations
- Evaluated only on three image benchmarks with a single ResNet-50 backbone.
- Performance still falls below the fully supervised method at higher partial rates.
Experimental setup
Datasets: Office31 OfficeHome ImageCLEF
Benchmarks: Office31 OfficeHome ImageCLEF
Metrics: mean average precision (mAP)
Baselines: CC RC PRODEN LWC PICO DPLL PaPi CDS CoDA Supervised
Builds on: PICO CoDA PRODEN partial label learning
Closest papers by meaning
- RoDA: Robust Domain Alignment for Cross-Domain Retrieval Against Label Noise (AAAI 2025 · similarity 0.91)
- DiCA: Disambiguated Contrastive Alignment for Cross-Modal Retrieval with Partial Labels (AAAI 2025 · similarity 0.87)
- Semantic Feature Learning for Universal Unsupervised Cross-Domain Retrieval (NeurIPS 2024 · similarity 0.86)
- Unsupervised Cross-Domain Image Retrieval via Prototypical Optimal Transport (AAAI 2024 · similarity 0.86)
- Retrieval Across Any Domains via Large-scale Pre-trained Model (ICML 2024 · similarity 0.85)
- Cross-Modal and Uni-Modal Soft-Label Alignment for Image-Text Retrieval (AAAI 2024 · similarity 0.85)
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.