Codes for Towards Unified Surgical Skill Assessment (CVPR 2021).
- Recommended Environment: Python 3.7, Cuda 10.1, PyTorch 1.6.0
- Install dependencies:
pip3 install -r requirements.txt.
- Complete the access form of the JIGSAWS dataset and get the permission.
- Download our processed data for JIGSAWS from Baidu Yun (PIN:sa67) or Google Drive.
- Unzip the files by
zip --fix data.zip --out data_full.zip && unzip data_full.zip. - Put the data into the parent directory of the codes.
- The data includes following sub-directories:
video_encoded : Surgical videos after pre-processing.
label : Ground truth scores of surgical skills.
feature_resnet101 : ImageNet-pretrained ResNet features with ten-crop augmentation (Visual Path Input).
kinematics_GT_14_1 : Kinematic data of the robotic surgical instruments (Tool Path Input).
time_val_1 : The sequences indicating task completion time (Proxy Path Input).
gesture_prediction : Surgical event preditions from MS-TCN models (Event Path Input).
As for the clinical dataset used in the paper, it might be released later if approved.
Simply run python3 main.py --config some_config_file.json .
The config files for our full model under the JIGSAWS 4-fold cross-validation setting are provided in the configs folder.
Trained models and Tensorboard logs will be saved in the result folder.
Our trained models and logs are provided in the pre_result folder.
@inproceedings{liu2021towards, title={Towards Unified Surgical Skill Assessment}, author={Liu, Daochang and Li, Qiyue and Jiang, Tingting and Wang, Yizhou and Miao, Rulin and Shan, Fei and Li, Ziyu}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, pages={9522--9531}, year={2021} }
MIT