This project implements person and vehicle Re-Identification (ReID) tasks, supporting training, testing, and online inference workflows:

• Dataset processing scripts, model code, pre-trained models, and evaluation tools.
• A complete pipeline for environment setup, data preparation (for both person and vehicle datasets), training, and testing.
• Support for multiple person ReID benchmarks (CUHK, Market, ICFG-PEDES, RSTP-Reid).
• Vehicle ReID datasets from real-road videos, including object extraction, data cleaning, and dataset splitting.

• [2025, 08]: We have released the complete ReID training and testing code.

git clone https://github.com/OPA067/ReID
cd ReID

conda create -n ReID python=3.12
conda activate ReID
pip install -r requirements.txt
pip install torch==1.8.1+cu102 torchvision==0.9.1+cu102 -f https://download.pytorch.org/whl/torch_stable.html

• 📂 reid_datasets
  • person_images
    • cam_a
    • cam_b
    • CUHK01
    • CUHK03
    • Market
    • ICFG-PEDES-train
    • ICFG-PEDES=test
    • train_query
    • test_query
    • RSTP-Reid
  • person_jsons
    • test_query.json
      • {"id": 0, "tar_path": "test_query/p3810/p3810_s4782.jpg", "can_path": "test_query/p3810/p3810_s4781.jpg"},
      • {"id": 1, "tar_path": "test_query/p7463/p7463_s9933.jpg", "can_path": "test_query/p7463/p7463_s9932.jpg"},
      • ...
    • train_cam_a_b.json
    • train_CUKH01.json
    • ...

where cam_a, cam_b, CUHK01, CUHK03, Market, train_query and test_query from CUHK-PEDES, RSTP-Reid from RSTP-Reid, and ICFG-PEDES-train and ICFG-PEDES-test from ICFG-PEDES, where .json stores the relative path of the image, e.g., "cam_a/010_0.bmp", and create .json details via:

python ./reid_datasets/create_person_reid_json.py and ./reid_datasets/create_car_reid_json.py

We collect real-road videos for object extraction, data cleaning, and train-test splitting, rather than using public datasets. The directory structure is as follows:

• 📂 reid_datasets
  • car_images
    • train_video_1
    • train_video_2
    • train_video_3
    • ...
    • test_video_1
    • test_video_2
    • test_video_3
    • ...
  • car_jsons
    • train_video_1.json
    • ...
    • test_video_1.json
    • ...

where data organization is identical to that of the person dataset, and note that train_video_1 and test_video_1 do not represent the same video.

The above models are stored as needed in the ./pretrain_models directory.

#!/bin/bash
root_dir={YOUR WORDING ROOT DIR}
DATASET_NAME=RSTP-Reid
CUDA_VISIBLE_DEVICES=0 \
    python retrieval_train.py \
    --name RDE \
    --img_aug \
    --txt_aug \
    --batch_size 32 \
    --root_dir $root_dir \
    --output_dir experiments \
    --dataset_name $DATASET_NAME \
    --loss_names ReID  \
    --pretrain_choice {PRETRAIN MODEL} \
    --log_period 100 \
    --num_epoch 20

or you can run:

bash reid_train.sh

python reid_test.py

To enable real-time visualization of the retrieval process without pre-storing all candidate images, we integrate the detection and retrieval processes, directly showing how the target image matches with all candidate images, as illustrated below:

python reid_online.py

Feature enhancement includes single-feature re-representation and multi-feature aggregation. Single-feature re-representation involves remapping the [CLS] token using methods such as MLP or Transformer, while multi-feature aggregation aggregates [Patch] tokens through learnable modules like MHA or Cluster. For these two types of feature enhancement schemes, the following feature alignment methods are proposed:

$S=sim({I}_1, {I}_2) = \frac{{I}_1 \cdot {I}_2}{||{I}_1||_2 \cdot ||{I}_2||_2}$

# update model/clip_model.py
def forward(self, tar_images, can_images):
    with torch.no_grad():
        tar_feats = self.encode_image(tar_images)
        can_feats = self.encode_image(can_images)

where $I$=[CLS]$\in \mathcal{R}^{1 \times D}$.

$S=sim({I}_1, {I}_2) = \frac{{I}_1 \cdot {I}_2}{||{I}_1||_2 \cdot ||{I}_2||_2}$

# update model/clip_model.py
def forward(self, tar_images, can_images):
    tar_feats = self.encode_image(tar_images)
    can_feats = self.encode_image(can_images)

$S = \frac{1}{2} \left( sim({I}_1, {I}_2) + sim({P}_1, {P}_2) \right ) = \frac{1}{2} \left( \frac{{I}_1 \cdot {I}_2}{||{I}_1||_2 \cdot ||{I}_2||_2} + \frac{{P}_1 \cdot {P}_2}{||{P}_1||_2 \cdot ||{P}_2||_2} \right)$,
where $I$=[CLS]$\in \mathcal{R}^{1 \times D}$ and $P$=[Patch]$=\frac{1}{N}\sum_i^N P_i\in \mathcal{R}^{1 \times D}$.

$S = \frac{1}{2} \left( sim({I}_1, {I}_2) + sim({P}_1, {P}_2) \right ) = \frac{1}{2} \left( \frac{{I}_1 \cdot {I}_2}{||{I}_1||_2 \cdot ||{I}_2||_2} + \frac{{P}_1 \cdot {P}_2}{||{P}_1||_2 \cdot ||{P}_2||_2} \right)$,
where $I$=[CLS]$\in \mathcal{R}^{1 \times D}$ and $P$=[Patch]$=Model(P) \in \mathcal{R}^{1 \times D}$. $Model$ can use MLP, MHA, or PTM.

行人再识别P2P技术报告1(初版) update 2025, 04.
行人再识别P2P技术报告2(优化) update 2025, 04.
行人再识别P2P技术报告3(补充) update 2025, 05.
PTM of ReID: Patch Token Merge update 2025, 05.
行人检索完整方案 update 2025, 06.
车辆再检索研究报告 update 2025, 07.
万物再检索研究报告 update 2025, 07.
轻量型密集型特征提取器实验报告 update 2025, 08.

If you encounter any issues, feel free to contact 223081200014@smail.swufe.edu.cn

Our code is based on CVPR2024RDE, CVPR2024HBI. We sincerely appreciate for their contributions.