Quantitative comparison of motion history image variants for video-based depression assessment

Abstract

Abstract Depression is the most prevalent mood disorder and a leading cause of disability worldwide. Automated video-based analyses may afford objective measures to support clinical judgments. In the present paper, categorical depression assessment is addressed by proposing a novel variant of the Motion History Image (MHI) which considers Gabor-inhibited filtered data instead of the original image. Classification results obtained with this method on the AVEC’14 dataset are compared to those derived using (a) an earlier MHI variant, the Landmark Motion History Image (LMHI), and (b) the original MHI. The different motion representations were tested in several combinations of appearance-based descriptors, as well as with the use of convolutional neural networks. The F1 score of 87.4% achieved in the proposed work outperformed previously reported approaches.

Keywords

• Depression assessment
• Affective computing
• Machine learning
• Image processing
• Facial image analysis
• Motion history image