Estimation of continuous valence and arousal levels from faces in naturalistic conditions, Nature Machine Intelligence 2021
Official implementation of the paper "Estimation of continuous valence and arousal levels from faces in naturalistic conditions", Antoine Toisoul, Jean Kossaifi, Adrian Bulat, Georgios Tzimiropoulos and Maja Pantic, published in Nature Machine Intelligence, January 2021 [1]. Work done in collaboration between Samsung AI Center Cambridge and Imperial College London.
Please find a full-text, view only, version of the paper here.
The full article is available on the Nature Machine Intelligence website.
[Demo] Discrete Emotion + Continuous Valence and Arousal levels | [Demo] Displaying Facial Landmarks |
---|---|
![]() |
![]() |
Automatic emotion analysis from faces in-the-wild
The code requires the following Python packages :
Pytorch (tested on version 1.2.0)
OpenCV (tested on version 4.1.0
skimage (tested on version 0.15.0)
We provide two pretrained models : one on 5 emotional classes and one on 8 classes. In addition to categorical emotions, both models also predict valence and arousal values as well as facial landmarks.
To evaluate the pretrained models on the cleaned test sets, simply run :
python test.py --nclass 8
where nclass defines which model you would like to test (5 or 8).
The program will output the following results :
Expression
ACC=0.82
Valence
CCC=0.90, PCC=0.90, RMSE=0.24, SAGR=0.85
Arousal
CCC=0.80, PCC=0.80, RMSE=0.24, SAGR=0.79
Expression
ACC=0.75
Valence
CCC=0.82, PCC=0.82, RMSE=0.29, SAGR=0.84
Arousal
CCC=0.75, PCC=0.75, RMSE=0.27, SAGR=0.80
If you use this code, please cite:
@article{toisoul2021estimation,
author = {Antoine Toisoul and Jean Kossaifi and Adrian Bulat and Georgios Tzimiropoulos and Maja Pantic},
title = {Estimation of continuous valence and arousal levels from faces in naturalistic conditions},
journal = {Nature Machine Intelligence},
year = {2021},
url = {https://www.nature.com/articles/s42256-020-00280-0}
}
[1] "Estimation of continuous valence and arousal levels from faces in naturalistic conditions", Antoine Toisoul, Jean Kossaifi, Adrian Bulat, Georgios Tzimiropoulos and Maja Pantic, published in Nature Machine Intelligence, January 2021
Code available under a Creative Commons Attribution-Non Commercial-No Derivatives 4.0 International Licence (CC BY-NC-ND) license.