Facial Action Unit Recognition With Multi-models Ensembling
This work addresses facial expression analysis for affective computing applications, but it is incremental as it combines existing methods without major innovations.
The paper tackled facial action unit recognition by ensembling multiple pre-trained models on the Aff-Wild2 dataset, achieving an F1 score (macro) of 0.731 on the validation set.
The Affective Behavior Analysis in-the-wild (ABAW) 2022 Competition gives Affective Computing a large promotion. In this paper, we present our method of AU challenge in this Competition. We use improved IResnet100 as backbone. Then we train AU dataset in Aff-Wild2 on three pertained models pretrained by our private au and expression dataset, and Glint360K respectively. Finally, we ensemble the results of our models. We achieved F1 score (macro) 0.731 on AU validation set.