search

XT5un / P2NeRF

[CVPR2024] Global and Hierarchical Geometry Consistency Priors for Few-shot NeRFs in Indoor Scenes
10 stars 1 forks source link

confusion about SCADE metrics #1

Closed leonwu0108 closed 2 months ago

leonwu0108 commented 3 months ago

Congratulations on your acceptance! I find that in the paper you point out

In the official implementation of SCADE, the rendering images are post-processed with ground truth. For fair comparisons, we just report the results without post-processing.

Could you please give a further instruction of how they did the post-processing or the location of the processing code? Thanks very much.

XT5un commented 2 months ago

The SCADE code has fixed this issue (https://github.com/mikacuy/scade/commit/23139b164461169e32121234696df88580c3d644). In previous versions, SCADE would calculate a least squares transformation between GT and pred before evaluating the image metrics. We used the previous version of the SCADE for our comparison and removed that transform to ensure a consistent comparison with other methods.

leonwu0108 commented 2 months ago

OK, I understand it now. Thank you!