Hierarchical Knowledge-Guided Reasoning for Text-based Person Re-identification (HKGR)

This repository provides the official implementation of the paper:
Hierarchical Knowledge-Guided Reasoning for Text-based Person Re-identification

---

📌 Overview

Text-based person re-identification (TIReID) aims to retrieve pedestrian images using natural language descriptions. Existing methods often struggle with fine-grained alignment due to random token masking in Masked Language Modeling (MLM). To address this, we propose HKGR, a framework that leverages hierarchical scene graph knowledge to guide token masking and enhance cross-modal reasoning. Key components include:

• Hierarchical Knowledge-Guided Reasoning (HKGR): Object-, attribute-, and relation-level masking guided by scene graphs.

• Multi-Grained Semantic Alignment (MGA): Coarse- and fine-grained alignment via token selection and similarity constraints.

  

---

📊 Results

Benchmark Performance

Dataset	R@1	R@5	R@10	mAP
CUHK-PEDES	75.21	90.22	94.31	67.28
ICFG-PEDES	65.29	81.17	86.22	39.40
RSTPReid	63.10	83.60	88.90	49.74

---

🛠️ Installation

1. Clone the repository:

git clone https://github.com/[YourUsername]/HKGR.git  
cd HKGR  

2. Install dependencies:

pip install -r requirements.txt  

Requirements:

• Python 3.8+
• PyTorch 1.12+
• Transformers 4.20+
• CUDA 11.3+

---

🏃 Usage

Data Preparation

1. Download datasets: CUHK-PEDES, ICFG-PEDES, RSTPReid.
2. Organize data into ./data/[dataset_name] following the structure:

├── train  
│   ├── images  
│   └── text  
├── val  
└── test  

Training

Take the CUHK_PEDES dataset for example:

sh ./options/run_cuhk_pedes.sh  

Key Arguments:

• --root_dir: Dataset path(CUHK-PEDES, ICFG-PEDES, RSTPReid).
• --mlm_type: Token masking type (default: 'obj+attr+rel`).
• --loss_names: loss function (default: sdm+mlm).

Evaluation

CUDA_VISIBLE_DEVICES=0 python test.py --config_file 'path_of_checkpoint/configs.yaml' 

---

🙏 Acknowledgements

Some components of this code implementation are adopted from CLIP, TextReID and TransReID. We sincerely appreciate for their contributions.

📜Contact

If you have any question, please feel free to contact us. E-mail: zengruigeng@nudt.edu.cn