Embedding Time Information in Facial Emotion Recognition using LMUs.
Keywords
Loading...
Authors
Issue Date
2020-07-01
Language
en
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Emotion recognition for everyday life is hard. Everyday life comes with multiple
challenges: it is continuous, has an abundance of background noise and has to deal
with the very subtle and shifting nature of emotion. This thesis investigates whether
these challenges can be addressed using Legendre Memory Units (LMUs). LMUs can
integrate time information into the data, allowing the data to code for actual
movements, instead of just a series of frames. For this research the "BAUM-1:
Spontaneous AudioVisual Face Database of A ective and Mental States"-database was
used. To lter out background noise, facial points were extracted from the frames.
These points were processed by the LMUs. On the output of the LMUs a classi er was
trained and evaluated. To incorporate the rebellious nature of emotion, the
Pleasure-Arousal-Dominance-model (PAD-model) was used. The PAD-model is
considered a promising way of modeling human emotion. The result was an improved
classi cation accuracy, especially for Multilayer Perceptrons. Additionally it resulted
in a framework that models emotional states in PAD-space using emotion recognition.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen