Research map / Label-efficient and robust learning / Domain adaptation and generalization
Doubly Contrastive Learning for Source-Free Domain Adaptive Person Search
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 source-free domain adaptive person search (SFDA-PS) and a Doubly Contrastive Learning (DCL) method that adapts a source person search model to unlabeled target domains via a mean-teacher framework plus relation-embedded (ReC) and memory-aided (MaC) contrastive losses.
Problem
Person search models generalize poorly to new domains; existing domain adaptive person search needs labeled source data, which is impractical due to privacy and storage/transmission costs. The paper targets adapting a source-trained person search model to an unlabeled target domain without source da
Approach
Combine a mean-teacher baseline with two task-oriented contrastive modules: ReC applies a relation graph over detection-head proposals to enforce consistency among proposals of the same person and separation across persons/classes, and MaC uses a hybrid memory bank (cluster centroids, outliers, hard proposals) to make Re-ID embeddings more discriminative on pseudo-labeled target data.
Key results
- On CUHK-SYSU, DCL with SeqNet reaches 79.6 mAP and 81.2 Top-1 from a PRW-trained source model, surpassing prior DAPS methods (DAPS 77.6 mAP) without source data.
- On PRW, DCL with SeqNet reaches 38.4 mAP and 82.5 Top-1, versus 29.8 mAP for source-only SeqNet.
- Ablation on CUHK-SYSU shows mean-teacher, ReC, and MaC are complementary, with the full model giving 79.6 mAP vs 54.8 for source only.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the mean-teacher framework (ablation) | SeqNet · CUHK-SYSU | mAP: 54.8 -> 65.3 | Helped |
| Add ReC to the mean-teacher framework (ablation) | SeqNet · CUHK-SYSU | mAP: 65.3 -> 65.8 | Helped |
| Add MaC to the mean-teacher framework (ablation) | SeqNet · CUHK-SYSU | mAP: 65.3 -> 79.2 | Helped |
| Combine ReC and MaC with the mean-teacher framework (ablation) | SeqNet · CUHK-SYSU | mAP: 65.3 -> 79.6 | Helped |
| Add RGN to model proposal relationships (ablation) | DCL on SeqNet · CUHK-SYSU | mAP: 77.9 -> 79.6 | Helped |
| Add hard-proposal embeddings to clustered embedding memory (ablation) | DCL on SeqNet · CUHK-SYSU | mAP: 77.9 -> 79.2 | Helped |
| Adapt GLCNet using DCL instead of testing the source model directly | GLCNet · CUHK-SYSU | mAP: 61.4 -> 83.5 | Helped |
| Adapt SeqNet using DCL instead of testing the source model directly | SeqNet · PRW | mAP: 29.8 -> 38.4 | Helped |
Surprising findings
- The DCL-adapted models outperform several fully supervised person search methods (e.g., OIM, CTXGraph) on PRW mAP despite lacking source data.
- The DCL-adapted NAE on PRW gets 38.2 mAP while the adapted GLCNet on PRW gets only 30.0 mAP, lower than the other two adapted models, suggesting the gains vary by base model.
Limitations
- Relies on pseudo labels from the source model, which are noisy under domain shift; noise can accumulate.
- Gains are uneven across base models; adapted GLCNet on PRW shows only 30.0 mAP versus 38.2 and 38.4 for the others.
Experimental setup
Datasets: CUHK-SYSU PRW
Benchmarks: CUHK-SYSU PRW
Metrics: mAP Top-1 Recall AP
Baselines: Source only DAPS FOUS DDAM OIM CTXGraph HOIM AlignPS PSTR COAT GLCNet
Builds on: Mean Teacher UMT DAPS SeqNet NAE
Closest papers by meaning
- Domain Generalizable Person Search Using Unreal Dataset (AAAI 2024 · similarity 0.89)
- Mixture of Submodules for Domain Adaptive Person Search (CVPR 2025 · similarity 0.88)
- Localization-Anchored Instance Discrimination for Domain Adaptive Person Search (AAAI 2026 · similarity 0.87)
- Unsupervised Domain Adaptive Person Search via Dual Self-Calibration (AAAI 2025 · similarity 0.87)
- Divide and Conquer: Hybrid Pre-training for Person Search (AAAI 2024 · similarity 0.85)
- DM-Adapter: Domain-Aware Mixture-of-Adapters for Text-Based Person Retrieval (AAAI 2025 · similarity 0.84)
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.