GTR+: Generative Retrieval for Unsupervised Text-Based Person Search

This is the official PyTorch implementation of our paper Generative Retrieval for Unsupervised Text-Based Person Search. The paper link will be released soon.

Highlights

We propose GTR+ for unsupervised text-based person search, removing the need for expensive human-annotated descriptions. GTR+ combines:

• a three-tier description generation framework for producing fine-grained and diverse pseudo texts;

• an adaptive confidence-weighted retrieval learning framework to alleviate noisy supervision;

• LargeFine-Person, a large-scale benchmark for unsupervised TBPS pre-training.

Updates

• [2026-3-20] Initial release of code.
• ...

Requirements

Our experiments are mainly conducted on NVIDIA L40 GPUs. The code should also run on other GPUs with sufficient memory.

More dependency details are provided in requirements.txt.

Quick Start

git clone ...
cd ...
conda create -n blip -y python=3.10
conda activate blip
pip install -r requirements.txt

Training/Evaluation

The following scripts provide an example for training and evaluation. Please modify the dataset paths and checkpoint paths in the scripts before running.

# Training
bash shell/train.sh

# Evaluation
bash shell/eval.sh

Prepare Datasets

Download the CUHK-PEDES dataset, ICFG-PEDES dataset and RSTPReid dataset.

dataset_root/
├── CUHK-PEDES/
│   ├── imgs/
│   │   ├── cam_a/
│   │   ├── cam_b/
│   │   └── ...
│   └── reid_raw.json
├── ICFG-PEDES/
│   ├── imgs/
│   │   ├── test/
│   │   └── train/
│   └── ICFG_PEDES.json
├── RSTPReid/
│   ├── imgs/
│   └── data_captions.json
└── LargeFine-Person/
    ├── imgs/
    ├── LargeFine_Person_qa.json
    ├── LargeFine_Person_com.json
    └── LargeFine_Person_sty.json

LargeFine-Person Dataset

Download our pre-training dataset LargeFine-Person

Unsupervised TBPS Results with BLIP as Baseline

CUHK-PEDES

Method	Baseline	Fine-tuning	R@1	R@5	R@10	mAP	Checkpoint
GTR	BLIP		47.53	68.23	75.91	42.91	/
GAAP	BLIP		47.64	67.79	76.08	41.28	/
MUMA	BLIP		59.52	77.79	84.65	52.75	/
GTR+	BLIP		61.35	79.35	85.75	55.75	Download
GTR+ (Pre-trained)	BLIP	✗	62.65	78.80	84.76	55.27	Download
GTR+ (Pre-trained)	BLIP	✓	64.65	80.72	86.78	58.67	Download

ICFG-PEDES

Method	Baseline	Fine-tuning	R@1	R@5	R@10	mAP	Checkpoint
GTR	BLIP		28.25	45.21	53.51	13.82	/
GAAP	BLIP		27.12	44.91	53.56	11.43	/
MUMA	BLIP		38.11	56.01	63.96	19.02	/
GTR+	BLIP		47.81	64.97	71.94	28.75	Download
GTR+ (Pre-trained)	BLIP	✗	47.53	64.32	71.39	25.38	Download
GTR+ (Pre-trained)	BLIP	✓	52.78	67.94	73.91	33.99	Download

RSTPReid

Method	Baseline	Fine-tuning	R@1	R@5	R@10	mAP	Checkpoint
GTR	BLIP		45.60	70.35	79.95	33.30	/
GAAP	BLIP		44.45	65.15	75.30	31.21	/
MUMA	BLIP		54.35	76.05	83.65	40.50	/
GTR+	BLIP		54.75	75.15	83.50	43.79	Download
GTR+ (Pre-trained)	BLIP	✗	52.00	74.05	82.35	38.72	Download
GTR+ (Pre-trained)	BLIP	✓	55.70	76.55	84.25	43.86	Download

Supervised TBPS Results with IRRA as Baseline

CUHK-PEDES

Method	Baseline	Fine-tuning	R@1	R@5	R@10	mAP	Checkpoint
GTR+	IRRA		59.44	78.54	85.22	54.11	Download
GTR+	IRRA	✓	77.13	90.82	94.49	68.37	Download

ICFG-PEDES

Method	Baseline	Fine-tuning	R@1	R@5	R@10	mAP	Checkpoint
GTR+	IRRA		43.77	60.77	68.05	22.30	Download
GTR+	IRRA	✓	67.80	82.81	87.66	41.00	Download

RSTPReid

Method	Baseline	Fine-tuning	R@1	R@5	R@10	mAP	Checkpoint
GTR+	IRRA		50.45	73.45	82.35	37.68	Download
GTR+	IRRA	✓	69.05	86.90	92.25	54.19	Download

More Examples

More qualitative examples of generated descriptions and retrieval results are shown below.

Citation

If you find this code useful for your research, please cite our paper.

coming soon