Skip to content

bbenligiray/greedy-face-features

Repository files navigation

Greedy Search for Descriptive Spatial Face Features

My implementation of the following paper:

Gacav, C.; Benligiray, B.; Topal, C., "Greedy search for descriptive spatial face features," International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017.

The original paper was implemented by Caner Gacav in C++. I reproduced the results in Python without referring to the original code. The original code used dlib for SVM, this one uses scikit-learn. The results seem to be about 2-4% higher than what we have reported in the paper, probably because of the difference in SVM parameters/implementation.

What is this?

  • Spatial features are derived from displacements of facial landmarks. They are a kind of geometric feature that can be used for facial expression recognition.
  • A large number of spatial features can be extracted from a face, but they are not all equally descriptive.
  • In the face expression recognition literature, geometric features are hand-picked, dimension-reduced or used as is, with the redundancy.
  • In this study, we use sequential forward selection to obtain a small subset of spatial features that describes the facial expressions well.
  • In the figure below, you see an example subset. The changes in the indicated vertical or horizontal distances are the selected spatial features.
  • The proposed method delivers 88.7% recognition accuracy in the CK+ dataset, which is the highest performance among the methods that only use geometric features.

How to run?

Entering the following to the terminal works for a clean install of Ubuntu 16.04. You will need Internet connection to download the required packages.

$ cd /path/to/file
$ chmod +x setup.sh
$ ./setup.sh
$ python main.py

Program Options

You will be asked if you want to extract the features from images, or load the ones I have provided. If you want to extract the features yourself, you need to download Emotion_labels.zip and extended-cohn-kanade-images.zip from the CK+ dataset page, and put them in the project directory.

Then, you will be asked if you want to run the feature selection, or use the provided example selection. Feature selection is not a deterministic process, so the final result will not be the same every time.