Closed mhayoz closed 1 year ago
Hi, thanks for your interest. Yes, you are right it's a small bug and I have fixed it now. I copied the code from cal_psnr()
but forgot that depth is one channel. Thankfully it does not affect the conclusions.
Thank you for your answer. Does that mean that the true RMSE is sqrt(3) times higher than reported in the paper?
Yes, I think so. However, it does not affect any conclusions we drew because all methods used the same code.
I see. Thank you for the clarification
I have a question regarding the implementations of PSNR and RMSE in src/trainer/utils.py. In line 368 the RMSE is computed as: rmse = (((a - b)2 mask).sum() / (mask_sum 3.0))0.5
Why is there a factor 3 in the division? I would expect this to be: rmse = (((a - b)*2 mask).sum() / (mask.sum()))**0.5
Am I missing something?