eye-contact-cnn
This repository provides a deep convolutional neural network model trained to detect moments of eye contact in egocentric view. The model was trained on over 4 millions of facial images of > 100 young individuals during natural social interactions, and achives an accuracy comaprable to that of trained clinical human annotators.
Libraries used in our experiment
- PyTorch 0.4.0
- opencv 4.0.0
- numpy 1.16.2
- PIL 5.3.0
- pandas 0.23.4
- dlib 19.13.0 (optional if you want live face detection)
To run
with a webcam
python demo.py
on a video file
python demo.py --video yourvideofile.avi
on our sample video
Try this if you don't want to use dlib's face and instead test with pre-detected faces.
Comment out the first line of demo.py "import dlib" if you didn't install dlib.
python demo.py --video demo_video.avi --face demo_face_detections.txt
Demo video has been downloaded from here. I used this face detector to generate the face detection file.
Flags
--face: Path to pre-processed face detection file of format [frame#, min_x, min_y, max_x, max_y]. If not specified, dlib's face detector will be used.-save_vis: Saves the output as an avi video file.-save_text: Saves the output as a text file (Format: [frame#, eye_contact_score]).-display_off: Turn off display window.- Hit 'q' to quit the program.
Notes
- Output eye contact score ranges [0, 1] and score above 0.9 is considered confident.
- To further improve the result, smoothing the output is encouraged as it can help removing outliers caused by eye blinks, motion blur etc.
Citation
Please cite this paper in any publications that make use of this software.
@article{chong2020,
title={Detection of eye contact with deep neural networks is as accurate as human experts},
url={osf.io/5a6m7},
DOI={10.31219/osf.io/5a6m7},
publisher={OSF Preprints},
author={Chong, Eunji and Clark-Whitney, Elysha and Southerland, Audrey and Stubbs, Elizabeth and Miller, Chanel and Ajodan, Eliana L and Silverman, Melanie R and Lord, Catherine and Rozga, Agata and Jones, Rebecca M and et al.},
year={2020}
}
Link to the paper: here