I am not sure when it started to happen but the recent versions of Reality capture are producing very noisy results.
We run a multi camera rig 238 cameras, different lens and camera configs, fully cross polarised lighting.
We were getting some very good scans and have done since building our rigs, now our scans have dropped in quality dramatically to the point we cannot use them in production as Agisoft is now producing better results and our clients obviously are used to a standard which we can no longer deliver with reality capture.
We have tried going back to the legacy 1 version of the depth maps and this also produces noise, so something has changed.
Others in the industry are also reporting this change.
We have tried running default settings, tweaking various alignment settings, changed from jpegs to tiffs, etc , if there is something we are missing it would be really good to know.
After more testing , we built a set of the same camera and lens configs around 50 from the 238 and this built nice and smooth, so i believe it might have something to do with how its handling ,multiple lens and cameras in a multi cam rig.
Hi, if you go to your account at My - Sign in - Capturing Reality, there are downloads to previous versions of Realitycapture. Please let us know if you find there is a regression when processing the same data.
When you drag and drop in images from the same camera with different lenses please let me know if you are seeing any differences in the focal prior?
are you grouping at all?
Seems to be matching correctly.
We have tried grouping, i will try again and also see if i can regress back a version or two.
Thanks
So i have ran grouping of the same data set from two different machines and both results are the same
This is an update to this issue. We recalibrated the entire rig, after multiple shots, changes to settings, the problem still persists, the resulting mesh is just noisy and full of artifacts, this happens across multiple machines.
We expect some surface noise across human skin but in its current state RC has become unusable for our pipeline. I am convinced this started to happen when the version 2 was implemented and going back to legacy does not resolve the issue.
Hi,
Please share the data with us if you can. We made a request on the ticket you have. It would be good to take a look. Did you try using older versions of RC, which you can download from your account on the CapturingReality website?
I have tried the oldest one i can find on the website.
I will have to create a dataset as most of the one we have currently are under NDA`s
Thanks
We ran some more tests, and have concluded that the High setting - which has always been used is potentially causing the problem, we tried a data set on normal and the scan is much more similar to previous assets.
Although it still contains errors/noise.
Hi again.
I will have a data set ready today for you to have a look at.
I found a version 1.2 that we tested and we are having similar results so it may not be the software but something else that is happening.
How do i upload the data sets to you?
Best regards
Hello, you should have the invitation for the data upload in your mail.
