Apparent Personality Prediction using Multimodal Residual Networks with 3D Convolution
Keywords
Loading...
Authors
Issue Date
2018-07-01
Language
nl
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
In this thesis we propose a 3D apparent personality prediction model as extension of the multimodal residual neural network used for first impression analysis by Güçlütürk et al. [1]. The original model was trained on audio-visual data from YouTube videos and predicts the Big Five personality traits of the people in the video. The auditory data and the visual data were randomly selected within a clip, and thus not synchronized. The novel contribution of this research is to study the effect of extending the visual information over multiple frames, and of synchronizing the two modalities on the performance of the model. The model architecture was adapted to include these changes, and several new models were trained. Each performed better than the baseline models trained on the same dataset. Moreover, we provide evidence that temporal information improves the performance. However, a different network architecture is needed to prove the effect of the synchronization.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen