I have been using XMP files without issues with 8 precalibrated machine vision cameras. So, I know how to create proper XMP files.
I’m now trying a new setup, with 7 Canon DSLR cameras, still computing my own calibration and generating XMP files.
However, now the reconstructions have a random rigid transform added up, mostly a translation.
Using F5 (Create reconstruction using input) multiple times results in reconstructions with quite large coordinate system differences.
This doesn’t happen with my other 8 camera setup. The reconstruction noise changes but surfaces remain in perfect registration.
Attached is a screen shot showing two subsequent F5 calls with prior pose & calibration locked.
Everything seems fine but the exported meshes have a 1 cm translation difference as can be seen in the meshlab screenshot.
The expected results would be near perfect alignment since I’m using the same locked XMP files for both reconstructions.
I repeated this experiment over and over, the differences can reach 5 cm.
Also the scene viewer inside RealityCapture has very different coordinate system for the two components. This doesn’t occur for my other XMP setup.
I attached the XMP files. I can send you a link for the full dataset (700 MB) in a private email.
Thanks,
Mathieu