Research map / Image generation / Text-to-image diffusion
Diverse Person: Customize Your Own Dataset for Text-Based 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 Diverse Person (DP), which edits existing text-based person search images with Stable Diffusion using reference attribute images and rewrites captions with ChatGPT, producing augmented training data that boosts IRRA, SRCF, and IVT on CUHK-PEDES, ICFG-PEDES, and RSTPReid.
Problem
Text-based person search lacks large real-world datasets because of pedestrian privacy risks and annotation costs, and virtual datasets have domain gaps and low text diversity.
Approach
Edit real pedestrian images and their captions instead of generating from scratch: attribute-conditioned diffusion inpainting/editing with reference clothing/accessory images, regularized by cross-attention-to-segmentation loss, plus LLM rewriting of captions to substitute the edited attributes.
Key results
- IRRA+DP reaches 75.66% Rank-1 on CUHK-PEDES, versus 73.38% for IRRA (+2.28 points).
- IRRA+DP reaches 65.61% Rank-1 on ICFG-PEDES, up 2.15 over IRRA's 63.46%.
- IRRA+DP reaches 62.48% Rank-1 on RSTPReid, up 2.28 over IRRA's 60.20%, and beats all compared methods there.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add Diverse Person data augmentation to IRRA | IRRA · CUHK-PEDES | Rank-1 accuracy: 2.28% | Helped |
| Add Diverse Person data augmentation to IRRA | IRRA · ICFG-PEDES | Rank-1 accuracy: 2.15% | Helped |
| Add Diverse Person data augmentation to IRRA | IRRA · RSTPReid | Rank-1 accuracy: 60.20 -> 62.48 | Helped |
| Add Diverse Person data augmentation to SRCF | SRCF · CUHK-PEDES | Rank-1 accuracy: 4.25% | Helped |
| Add Diverse Person data augmentation to IVT | IVT · CUHK-PEDES | Rank-1 accuracy: 4.82% | Helped |
Surprising findings
- Synthetic samples' t-SNE features are intermingled with real data, suggesting small synthetic-to-real gap.
Limitations
- Gains depend on availability of reference attribute images; diversity is lower when fewer are available.
- Reliance on segmentation (SAM) and dependency parsing for attribute-phrase matching may introduce errors.
- Evaluation limited to three existing datasets and three baseline frameworks.
Experimental setup
Datasets: CUHK-PEDES ICFG-PEDES RSTPReid
Benchmarks: CUHK-PEDES ICFG-PEDES RSTPReid
Metrics: Rank-1 Rank-5 Rank-10 mAP
Baselines: IRRA SRCF IVT CFine LGUR SSAN ViTAA CAIBC
Builds on: Stable Diffusion IRRA SRCF IVT delayed conditioning
Closest papers by meaning
- Automatic Synthetic Data and Fine-grained Adaptive Feature Alignment for Composed Person Retrieval (NeurIPS 2025 · similarity 0.85)
- Divide and Conquer: Hybrid Pre-training for Person Search (AAAI 2024 · similarity 0.85)
- Modeling Thousands of Human Annotators for Generalizable Text-to-Image Person Re-identification (CVPR 2025 · similarity 0.85)
- Exploring the Potential of Large Vision-Language Models for Unsupervised Text-Based Person Retrieval (AAAI 2025 · similarity 0.85)
- An Empirical Study of CLIP for Text-Based Person Search (AAAI 2024 · similarity 0.84)
- Adaptive Uncertainty-Based Learning for Text-Based Person Retrieval (AAAI 2024 · 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.