search

machanic / AU_R-CNN

The official implementation code of paper: "AU R-CNN：Encoding Expert Prior Knowledge into R-CNN for Action Unit Detection".
https://arxiv.org/abs/1812.05788
87 stars 24 forks source link

Acc #19

Open lilaczhou opened 3 years ago

lilaczhou commented 3 years ago

Hi Do you evaluate the accuracy?

Mahdidrm commented 3 years ago

Hi @lilaczhou I think they have noted: To best of our knowledge, the F1 score of BP4D (3-fold testing protocol) is 63%, this is state-of-the-art result.

lilaczhou commented 3 years ago

Except the F1 score, I want to know if you evaluate the accuracy score?

Mahdidrm commented 3 years ago

Except the F1 score, I want to know if you evaluate the accuracy score?

I got it.. no Idea ... I have several questions too, hope he answer...

machanic commented 3 years ago

All the related works report the average F1 score, rather than average. Because the accuracy is not fair in the case that the data distribution over different AU classes are imbalanced. However, F1 score considers this case.

machanic commented 3 years ago

You can read the experimental section of my paper to get all the F1 scores.

cug-lucifer commented 3 years ago

Execuse me, could you share me the BP4D database?

machanic commented 3 years ago

@cug-lucifer I have BP4D database, but it must sign an agreement with the creator of BP4D, otherwise it will have law issue. You can contact with BP4D creator.