Just wanted to put leave feedback here. I have been playing with photogrammetry software (WebODM, Mushroom, Metashape) and wanted to give RC a try. A bit of a bumpy road when trying to make it work, but I kept plugging away at it since I was impressed how fast it was on my 8 year old PC laptop (and the only one where RC could work since my other computers/workstations are Linux or MBP). I was very surprised how quickly it had generated a mesh in Preview mode.
Anyways, I faced two issues when trying out version 1.3:
In Alignment, when I ran the process, it said 1140/1143 images used. But when I looked through the list, I counted about 50 images that were crossed off. So I am not sure why it said 1140 instead of about 1090. It also only showed a single component and not two components for the images that were thrown out.
When I tried to reconstruct the mesh in normal detail, I received CUDA errors (such as error 13 and just “CUDA error” with no additional info). I tried a lot of things such as opening up RC with the shift key to return it back to a fresh install state; I also uninstalled my old CUDA drivers and updated to latest version 12.3, but still received the same errors.
I was almost about to give up on RC when I decided to try a previous version (1.2.2). After installing the previous version and firing it up, everything magically worked:
The Alignment miscount was now correct and broke my input into two components – one set with 1094 images and another set with 48 images.
Reconstruction in Normal Detail works and takes a little over an hour. Very impressed – I used Meshroom and running the same dataset on a 64-core/128-thread AMD Threadripper with 512 GB RAM took over 12 hrs (this is counting the whole flow from alignment to texturing… but still… my laptop would not have stood a chance with Meshroom).
Are you able to reproduce the issue with the alignment in a new instance with those images?
The minimum requirement to be able to use RealityCapture is CUDA capabilities 3.0+, but it is recommended to have at least 6.1+.
To be able to use reconstruction with your GPU, you will have to change the depth map algorithm version to version 1:
Good to know about the ability to select different versions – I wish the warning message had let me know about that since I lost a lot of time there. I went back to v1.3 to give it a spin on the Meshing and that now worked without throwing a CUDA error message.
However, the align still produces issues – the first time it produced these results with things being as described in the first post - it matched 1140/1143 - but when you look at the details, a lot more photos are crossed out (I got similar results both when running on an Azure Windows VM and on my laptop):
I then did what you suggested and went back to v1.3 and still weird behavior – but different outcome. It only matched 48/1143 images and all the other photos are missing:
So, I think I will stick at v1.2 for the time-being. Maybe it just has to do with this really old laptop. I tried v1.3 on my wife’s slightly newer laptop which has an Nvidia GTX 1060, and then there were no issues. So not sure why my laptop and the Azure VM produced bad results for the Align task.
Are there any reasons to use v1.3 to get that Version 2 for Meshing? I am happy with the results on v1.2 with Version 1 Meshing. Though I cannot continue to use my wife’s laptop with the 1060, I could install it in on my Linux Workstation which has a 4070 (Windows VM with throughput to the GPU) … if you think the V2 meshing uplift is really worth it.
Hi TiberiusTyrone,
in next release there will be a warning if the GPU is not sufficient to use Version 2 for meshing and it will be set to Version 1.
Regarding the alignment issue, is that happening for various datasets? Is it happening after application restart ( clear the cache manually as described here, and launch the application whilst holding down shift, and you should see the following options:
Then select “make like a clean install”)
Also, your alignment settings are not ideal. Can you try just pre-defined ones?
A bit random, but I decided to re-run the alignment again in the same project after the initial fail and then it worked. Strange. I meshed those results and it looks just as good as in v1.2. And despite using v1 meshing algorithm, it also seemed to run faster in v1.3.
I will try your method for fun and see what happens. And at some point, try to do things on the 4070 to get that speed boost.
Is there a reason my alignment settings are not ideal? Just been playing around to see what everything does – I will admit that turning up the preselector too high leads to the alignment creating too much noise and causes errors so I have turned that down to about .25 of the max features per image. I have gone through the tutorial, which has been helpful especially explaining what each argument does.
Would there be any further documents/user manual/rules of thumb for Reality Capture?
