Research map / Label-efficient and robust learning / Domain adaptation and generalization
Effective Comparative Prototype Hashing 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 Comparative Prototype Hashing (CPH), which builds a domain-shared unit hypersphere via prototype contrastive learning and maps it to Hamming space with relation-preservation and quantization losses for unsupervised domain adaptive image retrieval.
Problem
Unsupervised domain adaptive hashing for image retrieval performs poorly because existing methods apply generic domain adaptation to hashing or align domains in the low-dimensional Hamming space, losing semantic information.
Approach
Perform category-level prototype contrastive learning (alignment and uniformity of source/target class prototypes) on a domain-shared unit hypersphere, then map that space to hash codes with relation-preservation and quantization losses from both domains.
Key results
- On cross-domain retrieval, CPH improves average mAP over the second-best method by 19.29% on Office-Home and 13.85% on Office-31 (64 bits).
- On single-domain retrieval, average improvements are 8.96% (P→R, Office-Home), 3.05% (A→D, Office-31) and 8.87% (M→U, Digits).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| transfer source-domain knowledge (ablation) | CPH · Office-Home A→R | mAP: 4.51 -> 71.18 | Helped |
| add Prototype Contrastive Loss (ablation) | CPH · Office-Home A→R | mAP: 67.26 -> 71.18 | Helped |
| add Quantization Loss (ablation) | CPH · Office-Home A→R | mAP: 62.94 -> 71.18 | Helped |
| add Relation Preservation Loss (ablation) | CPH · Office-Home A→R | mAP: 4.53 -> 71.18 | Helped |
| preserve source-domain knowledge (ablation) | CPH · Office-Home A→R | mAP: 35.50 -> 71.18 | Helped |
Limitations
- Relies on source labels and pseudo-labels for the target domain, whose quality is not analyzed in depth
- Some baseline results are copied from prior papers rather than re-run
Experimental setup
Datasets: Office-Home Office-31 Digits (MNIST, USPS)
Benchmarks: Office-Home Office-31 Digits
Metrics: mAP Top-K precision curve precision-recall curve
Baselines: LSH ITQ DSH SGH GraphBit ITQ+ GTH-g GTH-h PWCF DAPH PEACE DANCE
Builds on: ProPos DAN DANN DANCE DAPH
Closest papers by meaning
- Domain Adaptive Hashing Retrieval via VLM Assisted Pseudo-Labeling and Dual Space Adaptation (NeurIPS 2025 · similarity 0.93)
- Deep Unsupervised Hashing via External Guidance (ICML 2025 · similarity 0.88)
- Statistical Model-driven Similarity Hashing: Bridging Modalities for Efficient Unsupervised Retrieval (AAAI 2025 · similarity 0.88)
- KALAHash: Knowledge-Anchored Low-Resource Adaptation for Deep Hashing (AAAI 2025 · similarity 0.88)
- PromptHash:Affinity-Prompted Collaborative Cross-Modal Learning for Adaptive Hashing Retrieval (CVPR 2025 · similarity 0.87)
- Lightweight Contrastive Distilled Hashing for Online Cross-modal Retrieval (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.