[BUG] Evaluation bug in eval.py compute_depth_metrics() about sharpnet

[MichaelRamamonjisoa]

Thank you! Changing from log10 to log would not change the ranking of different methods we evaluated, but numerical values could differ.
However, looking at the DORN paper and especially its evaluation code score_nyuv2.m (downloaded here ), they have the following:

C = abs(log10(A) - log10(B));
depth_score(7, i) = sum(C(:))/N;

Do you have examples of benchmarks natural log is used instead?

[huangxGo]

Thank you! Changing from log10 to log would not change the ranking of different methods we evaluated, but numerical values could differ.
However, looking at the DORN paper and especially its evaluation code score_nyuv2.m (downloaded here ), they have the following:

C = abs(log10(A) - log10(B));
depth_score(7, i) = sum(C(:))/N;

Do you have examples of benchmarks natural log is used instead?

Thank for your reply!I think you're right.I thought the RMSElog should be same as this line.

SharpNet/eval.py

Line 69 in 6d69dcd

diff_log = mask * (np.log(np.clip(input, a_min=1e-12, a_max=1e12)) - np.log(

It seems that it's not.As the paper and code you mentioned above, there is a little difference between 2 line of the evaluation code.And I will close this issue!