Below is a collection of metrics we applied for report facial action unit (AU) results. However, these metrics can be used for evaluating various applications involving object or event detection.

1. ROC curve and its area under the curve
2. Frame-based F1 score
3. Skew-normalized F1 score
4. Event-based F1-score

The project tree is organized as follows.

func/           directory for utility functions
README.md       this file
demoMet.m       demo of metrics given a ground truth label and a prediction

In Matlab shell:

>> cd metrics
>> demoMet

Given that the ground truth annotation label and predicted decision value decV are vectors of the same length , the function signature for different metrics are as follows:

1. ROC: metR = getROC(label, decV)
2. F1-frame: metF = getF1F(label, decV)
3. F1-norm: metN = getF1N(label, decV)
4. F1-event: metE = getF1E(label, decV)

The script has been tested on Windows 8, Ubuntu 3.11.0 and Mac X 10.9.4 without problems. Please send your feedbacks to Wen-Sheng Chu regarding any issues, bugs and improvements.

• Links: [ paper ]
• Contact: Please send comments to Wen-Sheng Chu ([email protected])
• Citation: If you use this code in your paper, please cite either of the following:

@inproceedings{ding2013facial,
  title={Facial Action Unit Event Detection by Cascade of Tasks.},
  author={Ding, X. and Chu, W.-S. and {De la Torre}, F. and Cohn, J. F. and Wang, Q.},
  booktitle={ICCV},
  year={2013}
}

@article{ding2016cascade,
  title={Cascade of Tasks for Facial Expression Analysis},
  author={Ding, X. and Chu, W.-S. and {De la Torre}, Fernando and Cohn, J. F. and Wang, Q.},
  journal={Image and Vision Computing},
  year={2016},
}

The code may be redistributed under BSD license. Please send your feedbacks to Wen-Sheng Chu. :)