search

NVlabs / neuralangelo

Official implementation of "Neuralangelo: High-Fidelity Neural Surface Reconstruction" (CVPR 2023)
https://research.nvidia.com/labs/dir/neuralangelo/
Other
4.31k stars 387 forks source link

Question about evaluation of PSNR. #155

Closed HLinChen closed 5 months ago

HLinChen commented 9 months ago

From your codes, I find you use all images for training on both tnt and dtu datasets. So what are the test images? Can you provide the indices of test images?

hbai98 commented 7 months ago

Same, I gain a PSNR of 35.15 when validating on dtu_scan24 which surpasses the reported value 30.64 a large margin.

mli0603 commented 5 months ago

Hi @HLinChen and @hbai98

We do not separate the training/test images when reporting PSNR. As mentioned in the paper, we use PSNR as a metric to assess the capacity to overfit the input images and capture details. I hope this helps!

ThumbmasWalker commented 3 months ago

Hi @mli0603,

If you do not separate the training/test images for PSNR, then why are the reported PSNR scores so much lower than those obtained by the model when overfitting to all training views? I'm trying to compare to your approach, but it's unclear to me exactly how the PSNR reported in the paper were computed.

Thanks!