Hmm. Accuracy for these kinds of tests would be most affected by OpenSfM upgrades, I think. So we could improve overall coverage and noise etc, but maybe change how we’re handling cameras and change global characteristics of the dataset that would show up here.
So, something to test for a start: how does the error change with change to camera lens type: you have it on auto (by default) for both tests. Which means in the old version it would likely choose perspective and in the new would use the upgraded brown-conrady.
I would test perspective, brown, and fisheye in both 2.0 and recent, and compare between the 6 variations.
Also, for changes to the handling of GCPs, I’m a little fuzzier, but deep in the internals of opensfm is where this gets set. There’s a file with parameters that you can change manually. I can’t recall any ODM flags for controlling this weighting, but OpenSfM weights GCPs lower than their usual accuracy.
Finally, make sure you have accuracy set globally using --gps-accuracy or on each image using DOP flags or similar so the accuracy of the GPS for the image exif gets weighted appropriately.