Depression recognition from facial videos: Preprocessing and scheduling choices hide the architectural contributions

Abstract

AbstractDeep learning models have been widely applied in video‐based depression detection. It is observed that the diversity of preprocessing, data augmentation, and optimization techniques makes it difficult to fairly compare model architectures. In this study, the typical ResNet‐50 model is enhanced by using specific face alignment methods, improved data augmentation, optimization, and scheduling techniques. The extensive experiments on two popular benchmark datasets (AVEC2013 and AVEC2014) obtained competitive results, compared to sophisticated spatio‐temporal models for single streams. Moreover, the score‐level fusion approach based on two texture streams outperformed the state‐of‐the‐art methods. It achieved mean square errors of 5.82 and 5.50 on AVEC2013 and AVEC2014, respectively. These findings suggest that the preprocessing and training configurations result in noticeable improvements, which have been originally attributed to the network architectures.