Is it possible to predict action units?

[tvaranka]

In the paper "Frame-level Prediction of Facial Expressions, Valence, Arousal and Action Units for Mobile Devices" predictions are made for AUs. I wonder if this is possible with the code provided here?

[av-savchenko]

Thanks for your question. Yes, the abaw5_expr_va_au_train.ipynb contains all the required information to reproduce our paper

[tvaranka]

Thanks for the quick answer! For inference, which model would you recommend for the best AU performance?

[av-savchenko]

Please, take a look at the above-mentioned ipynb file, there are several markdown comments after section "AU (Action Units)" that compare performance for several pre-trained models. I believe, enet_b0_8_best_vgaf works better

[tvaranka]

Thanks a lot!