Non-verbal Facial Action Units-based Automatic Depression Classification

TL;DR

This paper introduces an automatic depression classification method based on facial action units, combining short-term and clip-based analysis, achieving over 75% accuracy on a balanced dataset.

Contribution

The study presents a novel approach integrating Gaussian Mixture Models and rank pooling for depression detection from facial expressions, improving accuracy over existing methods.

Findings

Achieved over 75% classification accuracy on AVEC 2019 dataset.

Both GMM-based and rank pooling-based models achieved at least 70% accuracy.

Combining both models leverages complementary information for better predictions.

Abstract

Depression is a common mental disorder that causes people to experience depressed mood, loss of interest or pleasure, feelings of guilt or low self-worth. Traditional clinical depression diagnosis methods are subjective and time consuming. Since depression can be reflected by human facial expressions, We propose a non-verbal facial behavior-based automatic depression classification approach. In this paper, both short-term behavior-based and clip-based depression classification are constructed. The final clip-level decision of short-term behavior-based depression detection is yielded by averaging the predictions of all short-term behaviors while we modelling behaviors contained in all frames based on two Gaussian Mixture Models. To evaluate the proposed approaches, we select a gender balanced subset from AVEC 2019 depression corpus containing 30 participants. The experimental results show that our method achieved more than 75% depression classification accuracy, where both GMM-based clip-level depression modelling and rank pooling-based short-term depression behavior modelling achieved at least 70% classification accuracy. The result indicates that our approach can leverage complementary information from both systems to achieve promising depression predictions from facial behaviors.