DepthAnything / Depth-Anything-V2

[NeurIPS 2024] Depth Anything V2. A More Capable Foundation Model for Monocular Depth Estimation
https://depth-anything-v2.github.io
Apache License 2.0
3.86k stars 336 forks source link

Bad evaluation results. #58

Closed indu1ge closed 4 months ago

indu1ge commented 4 months ago

Hi,

Thank you for your great efforts in building such a foundation MDE model. I downloaded the ViT-L model and conducted evaluations on the NYU test set. However, the evaluation result I obtained shows an AbsRel of 0.098, which is significantly different from the reported 0.045 in the paper. Could you please provide the corresponding evaluation code?

Thank you for your assistance.

Xeladoes commented 2 months ago

https://github.com/DepthAnything/Depth-Anything-V2/issues/88