NICUface

NICUface: Robust Neonatal Face Detection in Complex NICU Scenes

Models

Model	Backbone	Weights
NICUface-RF	RetinaFace	Link:
NICUface-Y5F	YOLO5Face (YOLOV5l)	Link: https://drive.google.com/file/d/16-OEMYuAuaNyd3vUcmlw1aCWsQHtDAYt/view?usp=sharing
FaceOrientation	YOLO5Face (YOLOV5l)	Link: https://drive.google.com/file/d/1IvPdph3ghr6bVJp4HEbAMJZ1ZiID4sAM/view?usp=sharing

NICUface Training

Two models are presented here for neonatal face orientation: NICUface-RF based on RetinaFace, and NICUface-Y5F based on YOLO5Face. We strongly recommend users to try both models for their neonatal application given that they are highly complementary in detecting diverse complex NICU scenes.

NICUface-RF Training

NICUface-Y5F Training

1. Data preprocessing: Face orientation estimation to standardize images.

python3 face_orientation_estimation.py --image_directory

This will create a north_oriented_faces directory with the supplied images from image_directory with faces oriented North.

2. Prepare your data into a new neonate directory as:

data/neonate/
    train/
      images/
      label.txt
    val/
      images/
      label.txt

3. Download YOLO5face repo

https://github.com/deepcam-cn/yolov5-face.git 

4. Download NICUface repo

https://github.com/GreenCUBIC/NICUface.git

5. Download NICUface-Y5F weights (presented in table above)

6. Train model

python3 train_nicuface_y5f.py --data data/neonate.yaml --cfg models/model_nicuface_y5f.yaml --weights weights/nicuface_y5f.pt --hyp data/hyp.nicuface_y5f.yaml

Testing

After training, the resulting weights are saved under runs/train/exp1/last.pt and can be used in inference as:

test_nicuface_y5f.py --runs/train/exp1/last.pt 

NICUface Evaluation

Citation

Souley Dosso Y, Kyrollos D, Greenwood K, Harrold J, Green JR, 2022, "NICUface: Robust Neonatal Face Detection in Complex NICU Scenes," IEEE Access, 10:62893-62909, doi:10.1109/ACCESS.2022.3181167.

Link: https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9791241

Contact

• Yasmina Souley Dosso yasminasouleydosso@cmail.carleton.ca
• Daniel Kyrollos DanielKyrollos@cmail.carleton.ca
• James Green jrgreen@sce.carleton.ca