Research map / Label-efficient and robust learning / Domain adaptation and generalization
Unsupervised Domain Adaptive Person Search via Dual Self-Calibration
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 Dual Self-Calibration (DSCA), a UDA person search framework that filters noisy image-level pseudo-boxes with a perception-driven adaptive threshold and self-calibrating filter, and uses cluster-level proxy memory to limit instance pollution.
Problem
Unsupervised domain adaptive person search suffers from noisy pseudo-labels (low-quality pseudo bounding boxes from unsupervised detection and misidentified pseudo-IDs from clustering) caused by inter-domain gaps.
Approach
Combine an adaptive, perception-driven threshold and a self-calibrating high-order soft threshold filter applied at each backbone layer to suppress background/noisy boxes, with a cluster-level proxy memory (one proxy per cluster, updated by cluster) instead of instance-level memory, clustering once per epoch.
Key results
- On CUHK-SYSU as target (from PRW), DSCA reaches 80.2% mAP and 81.7% top-1, outperforming DDAM (79.5% / 81.3%) and DAPS (77.6% / 79.6%).
- On PRW as target (from CUHK-SYSU), DSCA reaches 39.9% mAP and 81.6% top-1, versus DDAM's 36.7% / 81.2%.
- Ablation shows removing PDAF and CPR together drops CUHK-SYSU mAP from 80.2 to 77.6.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DSCA instead of DDAM | UDA person search with ResNet-50 · PRW test | mAP: 36.7 -> 39.9 | Helped |
| Use DSCA instead of DDAM | UDA person search with ResNet-50 · PRW test | top-1: 81.2 -> 81.6 | Helped |
| Use DSCA instead of DDAM | UDA person search with ResNet-50 · CUHK-SYSU test | mAP: 79.5 -> 80.2 | Helped |
| Use DSCA instead of DDAM | UDA person search with ResNet-50 · CUHK-SYSU test | top-1: 81.3 -> 81.7 | Helped |
| Add PDT, replacing global average thresholding (ablation) | DSCA · PRW test | mAP: 39.0 -> 39.9 | Helped |
| Add SCF (ablation) | DSCA · PRW test | mAP: 38.4 -> 39.9 | Helped |
| Add both PDT and SCF (ablation) | DSCA · PRW test | mAP: 36.2 -> 39.9 | Helped |
| Add CPR (ablation) | DSCA · PRW test | mAP: 39.2 -> 39.9 | Helped |
| Replace AMT with PDT while using SCF (ablation) | DSCA PDAF · PRW test | mAP: 39.2 -> 39.9 | Helped |
| Replace Hard threshold with SCF while using PDT (ablation) | DSCA PDAF · PRW test | mAP: 38.1 -> 39.9 | Helped |
| Replace Soft threshold with SCF while using PDT (ablation) | DSCA PDAF · PRW test | mAP: 38.4 -> 39.9 | Helped |
| Replace PDT with AMT and SCF with Hard threshold (ablation) | DSCA PDAF · PRW test | mAP: 39.9 -> 38.6 | Hurt |
Surprising findings
- Removing the threshold (PDT) alone or the filter (SCF) alone degrades performance on both datasets, but the proposed hard/soft threshold combinations sometimes underperform the single-component variants on PRW.
Limitations
- Performance remains well below fully supervised SOTA methods (e.g., SEAS 97.1% mAP on CUHK-SYSU).
- Relies on DBSCAN clustering and pseudo-boxes from the source-pretrained detector, so pseudo-label quality is bounded by it.
- Training epoch counts are tuned per target dataset to avoid overfitting.
Experimental setup
Datasets: CUHK-SYSU PRW
Benchmarks: CUHK-SYSU PRW
Metrics: mAP top-1 accuracy
Baselines: DAPS DDAM OIM SeqNet PSTR SEAS DICL
Builds on: DAPS NAE SPCL DBSCAN
Closest papers by meaning
- Instance-Guided Scene Adaptation for Unsupervised Person Search (AAAI 2026 · similarity 0.92)
- Domain Generalizable Person Search Using Unreal Dataset (AAAI 2024 · similarity 0.88)
- Mixture of Submodules for Domain Adaptive Person Search (CVPR 2025 · similarity 0.87)
- Doubly Contrastive Learning for Source-Free Domain Adaptive Person Search (AAAI 2025 · similarity 0.87)
- Can We Evaluate Domain Adaptation Models Without Target-Domain Labels? (ICLR 2024 · similarity 0.86)
- Link-based Contrastive Learning for One-Shot Unsupervised Domain Adaptation (CVPR 2025 · 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.