EdgeTrans4Mark

pytorch implementation of ECCV2022 "One-Shot Medical Landmark Localization by Edge-Guided Transform and Noisy Landmark Refinement"

Dataset

Download image & annotations from IEEE ISBI2015 Challenge. Or use provided data copy DATA COPY(file extraction code: xjvh) with converted coco format annotations. Then place this dataset under /data folder.

${ROOT}
`-- data
    `-- cephalometric
`-- pretrained_models 
    `-- hrnetv2_w18_imagenet_pretrained.pth

Requirements

conda create -n landmark python==3.6.3
pip install -r requirements.txt

For torch and torchvision, you can find whl in pytorch_whl and pip install offline.

Training & Testing

• train stage1

CUDA_VISIBLE_DEVICES=0,1 python3 scripts/train_st1.py \
--cfg experiments/cephalometric/train_st1.yaml \
--gpus 0,1 

• use stage1 model to infer label

CUDA_VISIBLE_DEVICES=0 python3 scripts/test_st1.py \
--model [BEST STAGE1 MODEL] \
--cfg experiments/cephalometric/train_st1.yaml \
--gpus 0 --local-iter 4 --infer-train 

• train stage2

CUDA_VISIBLE_DEVICES=0,1 python3 scripts/train_st2.py \  --cfg experiments/cephalometric/train_st2.yaml \
--gpus 0,1 

Acknowledgements

Great thanks for the following works and their opensource codes HRNet, DETR.