Potentially more accurate depth map models to look at

[gituser123456789000]

https://diffusion-vision.github.io/dmd/

https://github.com/YvanYin/Metric3D
https://jugghm.github.io/Metric3Dv2/

Only looking at the maps and numbers provided as examples.. both look to have very nice accuracy. The second one provides numbers comparing itself to Depth-Anything

[gituser123456789000]

How do we run huggingface locally?

I saw Clone Repository was mentioned in another thread. I did that, but I don't know how to run it locally. I see the folder and the files locally, but how to run?

https://huggingface.co/spaces/JUGGHM/Metric3D

The output in the colored format ruins the depth map, because I can't convert that directly to the white front black rear standard, because both dark blue (front) and dark red (rear) both become the same shade (black), which is clearly run. The app.py file needs to be edited to output black and white, not colored

This may be the old version 1 by the way, not v2.. but still interesting to see, and I still need to learn how to run huggingface locally either way. Edit: The github says the huggingface is updated, so this very well may be v2)

[gituser123456789000]

https://jugghm.github.io/Metric3Dv2/resource_new/media/demo_full.mp4

[gituser123456789000]

I'm like 11 hours in to trying to run it locally, and trying to convert the colored maps to the correct white front / black rear format. I finally made breakthroughs and I'm progressing fast. I got results from the default v1 model that was in the original code, as well as the better new model.

I still need to get it to output the depth map in grayscale (white front / black rear).. and preferably to output only the depth map by itself instead of a combined image.. but I wanted to share the results. The accuracy is fantastic...

The Roman Colosseum's sky/clouds glithced.. but besides that, the rest looks excellent to me

[gituser123456789000]

I got some very nice results, but also very inconsistent results. I don't know if this is the model itself or the rest of the program that runs it. It seems to certainly have potential, but its current form isn't so easy to operate and get consistent results.. at least not for me.