MiDaS_Depth v1.0
This location is for Registered Users Only.
Perhaps you need to login or register.
13.1 or later
Linux, Mac, Windows
MiDaS_Depth is a nuke ML tool based on the MiDaS monocular depth estimation repo that can be found on github here: https://github.com/isl-org/MiDaS . It allows for the generation of a depth pass based on a single image with no camera or track required. It is not 100% temporally stable but can be frameblended or frameheld for use on image sequences. Changing the 'Input Resolution Scale' Slider can give much more detailed results but at the sacrifice of depth accuracy.
This tool is only compatible with nuke 13.1 due to the use of the new CatFileCreator node to create the .cat file. Due to the limitations of nuke 13.1 only being able to utilitze pytorch 1.6.0, we were only able to convert MiDaS v2.1 and not the very latest v3.0 model. However the largest model was used to give the highest resolutiuon and best results (sorry for the 400mb cat file). Also, the input colorspace should be SRGB but feel free to experiment with different colorspaces for different results.
Hopefully people can find uses for this tool; I have been experimenting using it to create point clouds that can then be deformed using Mads Hagbarth's Point Render tool to great effect!
Model File: https://drive.google.com/file/d/1F5D9eWh3gJzPs-vwUMC410t2xyeWo4PH
Comments
ERROR: Inference1: Exception caught processing model: The following operation failed in the TorchScript interpreter.
...
run.py(62): run
run.py(154):
RuntimeError: The size of tensor a (24) must match the size of tensor b (23) at non-singleton dimension 2
The full error is here:
https://pastebin.com/DGqm2LSH
Have you thought about incorporating this?
https://www.youtube.com/watch?v=lDeI17pHlqo
https://github.com/compphoto/BoostingMonocularDepth
RSS feed for comments to this post