Core Concepts
Novel lightweight models for facial emotion analysis in the ABAW competition show significant improvements in quality metrics.
Abstract
The article presents results from the sixth Affective Behavior Analysis in-the-wild (ABAW) competition focusing on facial expression recognition, valence-arousal estimation, and emotion intensity prediction. The study introduces lightweight models based on various architectures trained to recognize emotions from static photos. These models significantly improve quality metrics compared to existing techniques. The research emphasizes the importance of accurate emotion analysis for human-centered technologies and highlights challenges faced in unconstrained environments. The authors aim to construct single models that are fair, explainable, trustworthy, and privacy-conscious while achieving high performance in real-world scenarios.
Stats
Experimental results demonstrate significant improvements in quality metrics on validation sets compared to existing non-ensemble techniques.
MT-DDAMFN achieves top performance for VA estimation with a mean CCC 2% greater than the initial DDAMFN.
For EXPR classification and AU detection tasks, there are no significant gains using features from multi-task trained models.
Models like EmotiEffNet-B0 show high accuracy on AffectNet but may not perform as well across different datasets.