LDXray: A Dataset for X-ray Prohibited Items Detection

Introduction

LDXray is a large-scale dual-view dataset collected from real security scenes in transportation hubs. It contains a wide range of prohibited items categorized into 12 different classes.

Dataset Download

The dataset and code can be accessed from the following links:

• LDXray Dataset: Kaggle | BaiduNetDisk

Code

• Installation Please follow the official environment setup guide of MMdetection.

• Side-View pseudo-label Use ld-tools/sail2bbox.py to get pseudo-label for side-view data , you can define the target_category_ids to any classes you need.The input_json_path should be the main view annotation file.

• Main-View Slicing Use ld-tools/gen-main-view-slice-json-wider.py to get sliced main-view data for training sliced images detector.You can extract a subset of spesific classes first to get spesific classes dataset.

• Config Update the config file of dataset in ld-config/base/datasets.
  Update the num_classes in main config when taining specific classes of side-view model and sliced image model.

• Train

python tools/train.py ld-config/ld-retinanet/main-view-retinanet.py
python tools/train.py ld-config/ld-retinanet/sliced-retinanet.py
python tools/train.py ld-config/ld-retinanet/side-view-retinanet.py

• Side-view inference

python tools/test.py ld-config/ld-retinanet/side-view-retinanet.py work_dirs/side-view-retinanet/epoch_12.pth --out side-view-retinanet.pkl

• pkl2json Use ld-tools/pkl2jon.py to convert the side-view-retinanet.pkl to side-view-retinanet.pkl. The categoriy_id will be start from 1 ,you should update the categoriy_id by yourself.

• Side-view to main-view Use ld-tools/dv-proposal.py to get main-view region proposals from side-view prediction.

• Auxiliary Prediction Update the test config in sliced-retinanet.py(Using the new json of Step 'side-view to main-view' ) to inference main-view region proposals predicted by side-view model.

python tools/test.py ld-config/ld-retinanet/sliced-retinanet.py work_dirs/side-view-retinanet/epoch_12.pth --out sliced-retinanet.pkl

Citation

If you use LDXray in your research, please consider citing:

@inproceedings{tao2024dual,
  title={Dual-view X-ray Detection: Can AI Detect Prohibited Items from Dual-view X-ray Images like Humans?},
  author={Tao, Renshuai and Wang, Haoyu and Guo, Yuzhe and Chen, Hairong and Zhang, Li and Liu, Xianglong and Wei, Yunchao and Zhao, Yao},
  booktitle={2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2025},
  organization={IEEE}
}

Contact

For any questions, please contact Renshuai Tao.

---