Research map / Label-efficient and robust learning / Domain adaptation and generalization

Domain Generalizable Person Search Using Unreal Dataset

AAAI 2024 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Generalizable person re-identification · 4 citations (OpenAlex)

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 a domain-generalizable person search method trained only on the automatically labeled unreal JTA* dataset, using fidelity-adaptive training and domain-invariant feature learning to transfer to unseen real datasets (CUHK-SYSU, PRW).

Problem

Person search requires costly labeled real datasets and raises privacy issues; existing weakly-supervised and unsupervised domain adaptation methods still need target partial labels or retraining, so they do not generalize to arbitrary unseen datasets.

Approach

Train only on the automatically annotated unreal JTA* dataset, estimate per-instance fidelity (from BRISQUE-supervised deep features) to weight detection, confidence and ID-feature losses, and learn domain-invariant features via domain-guided normalization with a domain separation loss, treating each video sequence as a domain.

Key results

• Trained only on JTA*, the proposed method reaches 76.1 mAP / 78.4 Top-1 on CUHK-SYSU and 25.5 mAP / 79.4 Top-1 on PRW, versus baseline 66.7/71.0 and 20.9/76.0.
• Proposed method outperforms supervised baselines like HOIM, NAE, OIMNet++, SeqNet, and COAT in the JTA*-only DG comparison on CUHK-SYSU and PRW (e.g., 76.1 vs 66.3 mAP for OIMNet++ on CUHK-SYSU).
• On PRW, the proposed method (25.5 mAP) remains well below supervised SOTA such as AGWF (53.3 mAP) and COAT (53.3 mAP) that use the target training data.

Measured effects reported in the paper

Surprising findings

• Using all unreal instances without filtering degrades performance on real datasets, as degraded-visibility unreal instances hurt transfer.
• The proposed DG method is still well below supervised and weakly-supervised methods that train on target data on PRW (25.5 mAP vs ~50 mAP).

Limitations

• Performance on PRW remains substantially below methods that use target-domain training data.
• Only two real test datasets (CUHK-SYSU, PRW) are evaluated for generalization.
• Unreal dataset does not fully capture real-world styles, causing domain gap; occlusion filtering is heuristic (fewer than 13 occluded keypoints).

Experimental setup

Datasets: JTA* CUHK-SYSU PRW

Benchmarks: CUHK-SYSU PRW

Metrics: mAP Top-1 Precision Recall AP

Baselines: OIM HOIM NAE OIMNet++ SeqNet PSTR DMRNet++ COAT AGWF CGPS R-SiamNet CUCPS

Builds on: end-to-end person search network (OIM-style) adaptive gradient weighting function (AGWF) Faster R-CNN instance normalization BRISQUE

Closest papers by meaning

• Doubly Contrastive Learning for Source-Free Domain Adaptive Person Search (AAAI 2025 · similarity 0.89)
• Unsupervised Domain Adaptive Person Search via Dual Self-Calibration (AAAI 2025 · similarity 0.88)
• Unsupervised Multi-Source Domain Adaptation for Person Re-Identification (CVPR 2021 · similarity 0.87)
• Generalizable Person Re-Identification by Domain-Invariant Mapping Network (CVPR 2019 · similarity 0.87)
• Mixture of Submodules for Domain Adaptive Person Search (CVPR 2025 · similarity 0.85)
• Divide and Conquer: Hybrid Pre-training for Person Search (AAAI 2024 · similarity 0.85)