FSGS icon indicating copy to clipboard operation
FSGS copied to clipboard
verified publisher icon VITA-Group

Pearson Correlation Loss in paper or code

Open yanjh97 opened this issue 2 years ago • 4 comments

Hi, I found that the Pearson Correlation Loss of paper and code is inconsistent. in paper: Snipaste_2024-01-19_17-28-02 in code: https://github.com/VITA-Group/FSGS/blob/8c2e181821b6e21afe030aa52d155c7eac18eb47/train.py#L105-L108

yanjh97 avatar Jan 19 '24 09:01 yanjh97

Yes it makes me confused.

ShaohuaL avatar Apr 01 '24 12:04 ShaohuaL

I am also curious about this.

HkDzl avatar Apr 23 '24 07:04 HkDzl

The author has answered this question in issues/1

zhanghaoyu816 avatar Jul 01 '24 02:07 zhanghaoyu816

Hi guys, I'm still confused.

  1. the inverse depth means if the real depths is 50m, the inverse depth will be 1/50; but why - midas_depth in the code?
  2. the equation in the paper is: Corr(D_ras, D_est), which is implemented by pearson_corrcoef. The pearson_corr_coef is the same as the eq in the paper. image but the code is 1 - pearson_corrcoef. cc @zehaozhu @henrypearce4D @zhiwenfan

Master-cai avatar Aug 26 '24 02:08 Master-cai