Outdoor (VKITTI) model produces incorrect depth when max_depth=80, but works when max_depth=20

[az45uhuj]

Hi, Thanks for the great work on DepthAnything!

I am using the outdoor VKITTI pretrained metric model and notices incorrect / unstable depth predictions.

Problem description

When running inference with:

max_depth = 80.0

(the typical value for outdoor depth estimation), the model produces incorrect / unstable depth predictions.

However, when I reduce it to:

max_depth = 20.0

the predictions become reasonable and visually correct.

My question

How are the indoor and outdoor models trained regarding the depth range?

Does each pretrained model expect a specific max_depth value?

Should the user manually set max_depth, or should it be fixed per model?

Thanks in advance!

[honzadolezil]

Yes, each model expects the max_depth value it was trained with. You can take a look at the output of the DPT head, where there is sigmoid activation and then the whole output is rescaled using the max_depth parameter.