Depression Level Assessment based on 3D CNN and Facial Expression Videos

Abstract

Depression has a severe impact on people's daily lives and work, and it may even lead to suicide. Computer visionbased methods are promising for providing more effective and objective assistance in the clinical diagnosis of depression. In this article, to compare the performance of different 3D convolutional neural networks in assessing depression levels, we tested 3D VGGNet18, 3D GoogleNet, 3DEfficientNetB7, and 3D MobileNetV3 networks based on the AVEC2013 and AVEC2014 datasets. Experimental results showed that the 3D MobileNetV3 network achieved the best evaluation results, with MAE=7.35 and RMSE=9.16 on the AVEC2013 dataset, and MAE=7.19 and RMSE=9.08 on the AVEC2014 dataset. Compared with other existing methods, 3D ResNet18 demonstrated excellent performance.