Repository navigation
Outdoor (VKITTI) model produces incorrect depth when max_depth=80, but works when max_depth=20 #309
Description
Activity
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.
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:
(the typical value for outdoor depth estimation), the model produces incorrect / unstable depth predictions.
However, when I reduce it to:
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!