Greedy Search for Descriptive Spatial Face Features

Caner Gacav, Burak Benligiray, Cihan Topal

Published in: International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2017.

Read the paper

Get the source code

  • 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 can 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 uses geometric features.

Selected spatial features