search

nothingelse92 / AnchorFace

AnchorFace: An Anchor-based Facial Landmark Detector Across Large Poses(AAAI 2021)
66 stars 9 forks source link

Question about targeted confidence output #1

Closed DHdroid closed 3 years ago

DHdroid commented 3 years ago

Hello, I have a question about equation 5. image

Based on equation 9 (Pearson coefficient < 0), I believe that the targeted confidence should have negative relation with the L2 distance between the anchor pose and the ground-truth pose. If I am incorrect, would you mind providing me with an explanation as to why?

Thank you for your time and consideration :)

nothingelse92 commented 3 years ago

We sincerely apologize for the transcription error in Eq.5, it should be corrected as 截屏2021-02-05 下午9 06 30. We have revised our paper and uploaded it to the arxiv. Camera-ready version will be updated simultaneously. Thank you for your notice again, and our code will be released soon for reproduction.

DHdroid commented 3 years ago

Thank you for a kind reply. But I believe that f(x)=tanh^-1(x) is also increasing function wrt x, so I think it should be f(x)=tanh(x^-1). If I am incorrect, would you mind providing me with an explanation?

Thank you.

nothingelse92 commented 3 years ago

Actually we intended to use tanh^-1 (x) in Eq5 to represent inverse tanh (x), which is a decreasing function. We will rewrite the formula to make it more clear in our paper.