Affective and Implicit Tagging using Facial Expressions and Electroencephalography.
Abstract
Recent years have seen an explosion of user-generated, untagged multimedia data,
generating a need for efficient search and retrieval of this data. The predominant
method for content-based tagging is through manual annotation. Consequently, automatic
tagging is currently the subject of intensive research. However, it is clear that
the process will not be fully automated in the foreseeable future. We propose to involve
the user and investigate methods for implicit tagging, wherein users' responses
to the multimedia content are analysed in order to generate descriptive tags. We
approach this problem through the modalities of facial expressions and EEG signals.
We investigate tag validation and affective tagging using EEG signals. The former
relies on the detection of event-related potentials triggered in response to the presentation
of invalid tags alongside multimedia material. We demonstrate significant
differences in users' EEG responses for valid versus invalid tags, and present results
towards single-trial classification. For affective tagging, we propose methodologies to
map EEG signals onto the valence-arousal space and perform both binary classification as well as regression into this space. We apply these methods in a real-time
affective recommendation system.
We also investigate the analysis of facial expressions for implicit tagging. This
relies on a dynamic texture representation using non-rigid registration that we first
evaluate on the problem of facial action unit recognition. We present results on well-known
datasets (with both posed and spontaneous expressions) comparable to the
state of the art in the field.
Finally, we present a multi-modal approach that fuses both modalities for affective
tagging. We perform classification in the valence-arousal space based on these
modalities and present results for both feature-level and decision-level fusion. We
demonstrate improvement in the results when using both modalities, suggesting the
modalities contain complementary information.
Authors
Koelstra, Reinder Alexander LambertusCollections
- Theses [4495]