Modelling person-specific and multi-scale facial dynamics for automatic personality and depression analysis

Song, Siyang (2021) Modelling person-specific and multi-scale facial dynamics for automatic personality and depression analysis. PhD thesis, University of Nottingham.

[img]
Preview
PDF (Thesis - as examined) - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Available under Licence Creative Commons Attribution.
Download (30MB) | Preview

Abstract

‘To know oneself is true progress’. While one's identity is difficult to be fully described, a key part of it is one’s personality. Accurately understanding personality can benefit various aspects of human's life. There is convergent evidence suggesting that personality traits are marked by non-verbal facial expressions of emotions, which in theory means that automatic personality assessment is possible from facial behaviours. Thus, this thesis aims to develop video-based automatic personality analysis approaches. Specifically, two video-level dynamic facial behaviour representations are proposed for automatic personality traits estimation, namely person-specific representation and spectral representation, which focus on addressing three issues that have been frequently occurred in existing automatic personality analysis approaches: 1. attempting to use super short video segments or even a single frame to infer personality traits; 2. lack of proper way to retain multi-scale long-term temporal information; 3. lack of methods to encode person-specific facial dynamics that are relatively stable over time but differ across individuals.

This thesis starts with extending the dynamic image algorithm to modeling preceding and succeeding short-term face dynamics of each frame in a video, which achieved good performance in estimating valence/arousal intensities, showing good dynamic encoding ability of such dynamic representation. This thesis then proposes a novel Rank Loss, aiming to train a network that produces similar dynamic representation per-frame but only from a still image. This way, the network can learn generic facial dynamics from unlabelled face videos in a self-supervised manner. Based on such an approach, the person-specific representation encoding approach is proposed. It firstly freezes the well-trained generic network, and incorporates a set of intermediate filters, which are trained again but with only person-specific videos based on the same self-supervised learning approach. As a result, the learned filters' weights are person-specific, and can be concatenated as a 1-D video-level person-specific representation. Meanwhile, this thesis also proposes a spectral analysis approach to retain multi-scale video-level facial dynamics. This approach uses automatically detected human behaviour primitives as the low-dimensional descriptor for each frame, and converts long and variable-length time-series behaviour signals to small and length-independent spectral representations to represent video-level multi-scale temporal dynamics of expressive behaviours. Consequently, the combination of two representations, which contains not only multi-scale video-level facial dynamics but also person-specific video-level facial dynamics, can be applied to automatic personality estimation.

This thesis conducts a series of experiments to validate the proposed approaches: 1. the arousal/valence intensity estimation is conducted on both a controlled face video dataset (SEMAINE) and a wild face video dataset (Affwild-2), to evaluate the dynamic encoding capability of the proposed Rank Loss; 2. the proposed automatic personality traits recognition systems (spectral representation and person-specific representation) are evaluated on face video datasets that labelled with either 'Big-Five' apparent personality traits (ChaLearn) or self-reported personality traits (VHQ); 3. the depression studies are also evaluated on the VHQ dataset that is labelled with PHQ-9 depression scores. The experimental results on automatic personality traits and depression severity estimation tasks show the person-specific representation's good performance in personality task and spectral vector's superior performance in depression task. In particular, the proposed person-specific approach achieved a similar performance to the state-of-the-art method in apparent personality traits recognition task and achieved at least 15% PCC improvements over other approaches in self-reported personality traits recognition task. Meanwhile, the proposed spectral representation shows better performance than the person-specific approach in depression severity estimation task. In addition, this thesis also found that adding personality traits labels/predictions into behaviour descriptors improved depression severity estimation results.

Item Type: Thesis (University of Nottingham only) (PhD)
Supervisors: Valstar, Michel
Johnston, Alan
Keywords: machine learning, automatic personality analysis, computer vision
Subjects: Q Science > QA Mathematics > QA 75 Electronic computers. Computer science
Faculties/Schools: UK Campuses > Faculty of Science > School of Computer Science
Item ID: 65713
Depositing User: Song, Siyang
Date Deposited: 04 Aug 2021 04:43
Last Modified: 04 Aug 2021 04:43
URI: http://eprints.nottingham.ac.uk/id/eprint/65713

Actions (Archive Staff Only)

Edit View Edit View