The performance of DISFA dataset

[tghim]

The performance of DISFA dataset

[ZhiwenShao]

Thanks for your question. Here are the responses:

1. The DISFA model is trained from the best BP4D model. In our experiments, we use the model trained on the combination of BP4D part1 and part3. We do not use the whold BP4D dataset.
2. The results of our trained DISFA models (3-fold cross-validation) are:
   f1_score accuracy mean_error failure_rate
   0.577556 0.929593 0.043427 0.000000
   0.656207 0.939148 0.044319 0.000099
   0.671716 0.950716 0.042331 0.000162
   ==============================
   0.635159666666667 0.939819000000000 0.0433590000000000 8.70000000000000e-05
3. The local f1 score is generally worse than f1 score.

[tghim]

Thank you for your answer!
It helps me a lot