Problem with reproducing the results

[JacekMaksymiuk]

Hi, I wanted to reproduce your results, but I was unsuccessful.
I wanted to reproduce the results from your paper, but for the football dataset and the Depth-Anything-V2-Large model I get completely different results.
My results:
abs_rel: 0.6578764319419861
rmse: 0.3091728091239929
rmse_log: 1.8111008405685425
silog: 79.3974380493164
sq_rel: 0.2289116233587265

I also tried a bit of postprocessing the data (which is in float32 format), i.e. normalising it on the interval 0-65535, so as to lose as little information as possible, but this helped slightly:
abs_rel: 0.5723931789398193
rmse: 0.25019174814224243
rmse_log: 1.7408266067504883
silog: 71.5643539428711
sq_rel: 0.18551835417747498

Should I apply any more postprocessing? I know that scaling and shifting is done during evaluation, but as we can see this is not enough. How can I reproduce the results from your paper?

Thank you and best regards :)

[JacekMaksymiuk]

After applying simple transformations (normalizing the result between min(gt) and max(gt)), I was able to go down for the silog statistic from over 70 to about 1.95, but this is still a far from the paper's results.
Have you applied any other transformations on the result besides those implemented in evaluate.py?