Our technical report is available at https://arxiv.org/pdf/2309.17143.pdf

The online results on the public leaderboards could be viewed at board1 and board2. (Algorithm Name: SUTD-VLG, User Name: fivethousand).

Our method achieves 1st place ranking on three metrics and 3rd place on the remaining one.

To build a compatible conda env, you only need to run the following lines one by one:

conda create -n LMD python=3.10
conda activate LMD
pip install -r requirements.txt
pip install -U openmim
cd mmpose_package/mmpose
pip install -e .
mim install mmengine
mim install "mmcv>=2.0.0"
pip install --upgrade numpy

To validate the effectiveness of the built conda env, you could run step1_test_mmpose.py. It will report the version of the installed mmpose package.

As we do not have the right to forward the CLdetection dataset, interested researhers should follow this website to request it. After being approved, you will have access to 2 files: train_stack.mha and train-gt.json. Then you could download them and place them under the ./data. After that , you will have the following data structure under ./data:

.
├── train-gt.json
└── train_stack.mha

To make the dataset structure compatible with the MMPose package, you should convert the original dataset into a coco-style one with the provided script step2_prepare_coco_dataset.py.

python step2_prepare_coco_dataset.py --mha_file_path ./data/train_stack.mha --train_gt_path ./data/train-gt.json --image_save_dir ./data/preprocessed

It will generate the preprocessed dataset, together with the train.json, valid.json, test.json. Then, the ./data directory will have the following file structure:

.
├── preprocessed
├── test.json
├── train-gt.json
├── train.json
├── train_stack.mha
└── valid.json

Specify the config and the working directory, and run:

python step3_train_and_evaluation.py --config 'configs/CLdetection2023/srpose_s2.py' --work-dir './MMPose-checkpoints/CLdetection/srpose_s2'

Note that both the checkpoints and logs will be saved in the --work-dir.

Test with the pretrained weights:

python step4_test_and_visualize.py --config 'configs/CLdetection2023/srpose_s2.py' --checkpoint './MMPose-checkpoints/CLdetection/srpose_s2/demo.pth'

please refer to inference_single_image.py

python inference_single_image.py --config 'configs/CLdetection2023/srpose_s2.py' --checkpoint './MMPose-checkpoints/CLdetection/srpose_s2/demo.pth' --mode "val"

A model pre-trained on the combined train and val datasets is available at https://drive.google.com/file/d/10HrNDBBpuECTcgNgcWNUt7kM6m4ZMXM4/view?usp=sharing

Note that we ensembled 7 models for our final submission in the MICCAI CLdetection2023 challenge. You can try our submitted docker at https://cl-detection2023.grand-challenge.org/evaluation/challenge/leaderboard/ after applying for it. (You need to register an account on that platform and request access to the algorithm).

If you find this repo useful, please consider giving us a star and citing our paper.

@misc{wu2023revisiting,
      title={Revisiting Cephalometric Landmark Detection from the view of Human Pose Estimation with Lightweight Super-Resolution Head}, 
      author={Qian Wu and Si Yong Yeo and Yufei Chen and Jun Liu},
      year={2023},
      eprint={2309.17143},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

We would like to thank the MICCAI CLDetection2023 organizers for providing well-established baselines and their altruistic service for the contest. We appreciate all the contributors of the MMPose Package. We thank the authors of SRPose for making their code public.