Unable to Reproduce NYUv2 and KITTI Results Using TorchHub Models (Hugging Face Versions Work)

[ahmedshahabaz]

Hi,
Apologies in advance if I’m missing something, but I’ve been trying to reproduce the zero-shot relative depth estimation results on the NYUv2 and KITTI datasets using the MiDaS v3.1 models. When using the pre-trained models from the official GitHub or TorchHub (intel-isl/MiDaS), I wasn’t able to match the reported performance.

However, when I used the same models from the Hugging Face Hub (e.g., Intel/dpt-beit-large-384), I was able to recreate the expected results on both datasets.

Could there be a mismatch or issue with the weights provided via TorchHub or the pre/post-processing steps?

Thanks for your great work and support!