EDS dataset and the code implementation of PSN

CVPR 2022 paper 《Exploring Endogenous Shift for Cross-domain Detection: A Large-scale Benchmark and Perturbation Suppression Network》

English | 简体中文

Download Link of EDS Dataset:

Please follow the instructions in: https://github.com/DIG-Beihang/XrayDetection to obtain the EDS download link.

Cross-domain Transfer Learning

A key feature of this repository is its support for cross-domain transfer learning across different X-ray imaging domains. Users can conveniently specify different source and target domains through command-line arguments such as --dataset and --datasets, without manually modifying the training pipeline.

For example, a model can be trained on one X-ray domain and transferred to another domain with different imaging characteristics, device types, or data distributions:

python train.py \
    --dataset domain1 \
    --datasets domain2

By changing the values of --dataset and --datasets, users can flexibly construct different source-to-target transfer settings, such as domain1 → domain2, domain1 → domain3, or other cross-domain combinations supported by the dataset.

Different domains may exhibit substantial distribution shifts caused by different X-ray machines, imaging parameters, object appearances, and acquisition environments, as illustrated below:

Example of domain shifts across different X-ray imaging machines.

This design enables users to conveniently evaluate and adapt detection models under different domain settings, providing flexible support for studying cross-domain transferability, domain adaptation, and model generalization in practical X-ray detection scenarios.

Prerequisites

• Python 3.6
• Pytorch 0.4.1
• CUDA 8.0 or higher

Compile

pip install -r requirements.txt
cd lib
sh make.sh

Training

The scripts folder has all the training scripts. For example, if you want to train an experiment from domain1 to domain2, just run:

sh scripts/train-1-2-fc.sh

Testing

The scripts folder has all the testing scripts. For example, if you want to test a model trained from domain1 to domain2, just run:

sh scripts/test-all-1-2.sh

Citation

If this work helps your research, please cite the following paper.

@inproceedings{Tao:CVPR22,
  author    = {Renshuai Tao and Hainan Li and Tianbo Wang and Yanlu Wei and Yifu Ding and Bowei Jin and Hongping Zhi and Xianglong Liu and Aishan Liu},
  title     = {Exploring Endogenous Shift for Cross-domain Detection: A Large-scale Benchmark and Perturbation Suppression Network},
  booktitle = {IEEE CVPR},
  year      = {2022},
}