I feel like this is not mentioned anywhere, but it does seem to be a big problem if flying the drone non-rotated, I recently encountered a couple of datasets obtained this way, and I can’t get them constructed in ODM, colmap, and agisoft no matter how I adjusted the processing parameter. This is probably not a problem for people using flight plan software, but those who want to do a quick survey manually might need to take care.
This was mentioned by @YanNoun once Dense pointcloud boundary shrinks comparing to opensfm's sparse point cloud - #13 by YanNoun. I guess that the SFM algorithm cannot deal with some ambiguity when reconstructing with pure parallel cameras.
I’m also very curious about the math behind this problem but don’t know where to look into it. It would be appreciated if anyone who knows about this could give some explanation or simply some topics that I can search for.