Research map / Label-efficient and robust learning / Recognition and long-tail
Optimizing Label Assignment for Weakly Supervised 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.
Proposes Optimizing Label Assignment (OLA) for weakly supervised person search, combining graph-constrained clustering, optimal-transport prototype matching, and a dual memory bank to improve pseudo-label quality.
Problem
Weakly supervised person search (only bounding-box annotations) suffers from inaccurate pseudo labels from clustering at instance level and from MaxIoU-based proposal label assignment, plus background noise in the memory bank.
Approach
Adds bipartite-graph (Kuhn-Munkres) matching constraints into DBSCAN so two people in one scene cannot share an identity, assigns proposal labels via optimal transport over an IoU+cosine cost matrix, and separates instance and proposal memory banks linked by cross-attention.
Key results
- On CUHK-SYSU, OLA reaches 87.8% mAP and 89.3% Top-1, the best among weakly supervised methods in the paper's table.
- On PRW, OLA reaches 38.1% mAP and 82.0% Top-1 (text states 38.2% mAP and 83.3% Top-1, inconsistent with Table 2), exceeding DICL at 35.5% mAP.
- Ablation on PRW: adding all three components raises the baseline from 35.4% to 38.1% mAP.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OLA instead of DICL | weakly supervised person search · CUHK-SYSU test set, gallery size 100 | mAP: 87.4 -> 87.8 | Helped |
| Use OLA instead of DICL | weakly supervised person search · CUHK-SYSU test set, gallery size 100 | Top-1: 88.8 -> 89.3 | Helped |
| Use OLA instead of DICL | weakly supervised person search · PRW test set | mAP: 35.5 -> 38.1 | Helped |
| Use OLA instead of DICL | weakly supervised person search · PRW test set | Top-1: 80.9 -> 82.0 | Helped |
| Add Prototype Matching (ablation) | weakly supervised person search baseline · PRW | ReID mAP: 35.4 -> 36.4 | Helped |
| Add Prototype Matching (ablation) | weakly supervised person search baseline · PRW | ReID Top-1: 81.1 -> 81.5 | Helped |
| Add Prototype Matching (ablation) | weakly supervised person search baseline · PRW | Detection Recall: 62.1 -> 64.6 | Helped |
| Add Context-Aware Clustering after Prototype Matching (ablation) | weakly supervised person search · PRW | ReID mAP: 36.4 -> 37.2 | Helped |
| Add Context-Aware Clustering after Prototype Matching (ablation) | weakly supervised person search · PRW | ReID Top-1: 81.5 -> 81.7 | Helped |
| Add Dual Memory Bank Enhancement after Prototype Matching and Context-Aware Clustering (ablation) | weakly supervised person search · PRW | ReID mAP: 37.2 -> 38.1 | Helped |
| Add Dual Memory Bank Enhancement after Prototype Matching and Context-Aware Clustering (ablation) | weakly supervised person search · PRW | ReID Top-1: 81.7 -> 82.0 | Helped |
Limitations
- Reported PRW numbers are inconsistent between text (38.2% mAP, 83.3% Top-1) and Table 2 (38.1%, 82.0%).
- Relies on a cosine threshold of 0.5 for graph edges and hand-set hyperparameters (beta=0.2, k=3, alpha).
- Gains over prior weakly supervised methods are modest on PRW relative to fully supervised methods.
Experimental setup
Datasets: CUHK-SYSU PRW
Benchmarks: CUHK-SYSU PRW
Metrics: mAP Top-1 accuracy recall
Baselines: CGPS R-SiamNet SSL DICL DPM OIM SeqNet COAT
Builds on: SPCL OTA DBSCAN Faster R-CNN SeqNet
Closest papers by meaning
- Exploring Visual Context for Weakly Supervised Person Search (AAAI 2022 · similarity 0.90)
- Localization-Anchored Instance Discrimination for Domain Adaptive Person Search (AAAI 2026 · similarity 0.84)
- Hierarchical Online Instance Matching for Person Search (AAAI 2020 · similarity 0.84)
- WISH: Weakly Supervised Instance Segmentation using Heterogeneous Labels (CVPR 2025 · similarity 0.83)
- Divide and Conquer: Hybrid Pre-training for Person Search (AAAI 2024 · similarity 0.83)
- Relational Matching for Weakly Semi-Supervised Oriented Object Detection (CVPR 2024 · similarity 0.82)
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.