OPK Parser exception with DJI image

I try to use opensfm to parse images taken by a DJI drone equipped with RTK

To verify the correctness of the parser, I use https://github.com/mapillary/OpenSfM/pull/838#issuecomment-1212630210 mentioned method directly triangulation and obtains the sparse model and camera pose with poor accuracy.

But I found that, as shown in the above visualization, some images are completely wrong in the parsed orientation, while others are correct.

I used incremental reconstruction to obtain GT and tried to find the difference between the two, but I found that it was difficult to find a pattern. Some orientations differed by 180 degrees, and some differed by 90 degrees. How can I solve this problem?

Here is a set of images I used for testing: https://drive.google.com/file/d/15ETwKP3Hn1MgdJDdNs6-4ae5uH33xV9V/view?usp=drive_link

test method:

1. Perform an incremental rebuild using opensfm and observe that everything is OK
2. run ./bin/opensfm reconstruct --algorithm triangulation data/dji, Camera pose error observed

If the camera pose is wrong, then the pairs selection mechanism based on the opk parsing result is wrong, which is a bug for the system.

Can someone help me find the problem?

Download the linked DJI image set and reproduce the issue with ./bin/opensfm reconstruct --algorithm triangulation data/dji. Compare its camera poses and pair selection with the incremental reconstruction and inspect the OPK parsing path referenced by PR #838. Done means identifying and documenting the cause of the inconsistent 90- or 180-degree orientations, with a verified correction.