Geometrical, Graphical, and Transformer-based Approaches for Recognizing Naturalistic Depression in Facial Behavior

Depression is a serious mental illness that profoundly impacts the daily lives and well-being of individuals. In this context, smartphone cameras and sensors offer a ground-breaking opportunity for automated depression detection, as they enable real-time monitoring and large-scale data collection, providing valuable insights into the condition. Within the ABC Challenge 2025 framework, the FacePsy dataset, consisting of facial feature videos recorded in the wild, is utilized to tackle the challenge of naturalistic depression recognition. In this study, the DP2K team’s proposed pipeline leverages angle-based, distance-based, and deep graphical features extracted from facial landmarks and action units. Specifically, enhanced deep graphical features are extracted using a transformer-based model, which processes video summarization tokens from each frame. These features are then used for classification with a residual feedforward network. Additionally, XGBoost, LightGBM, and Random Forest are trained alongside the deep learning model for comparison. The models are evaluated in two scenarios: future forecasting and universal recognition, where the proposed model achieves $87.54 \%$ and $48.39 \%$ AUROC respectively. The results demonstrate the model’s outstanding performance, effectively balancing both scenarios. These findings underscore the model’s potential versatility for real-world applications, offering an innovative and scalable approach for ubiquitous depression detection.CCS Concepts: · Human-centered computing $\rightarrow$ Ubiquitous and mobile computing · Computing methodologies $\rightarrow$ Neural networks.