Use prior poses without any keypoints

[cduguet]

trafficstars

Hi!

I have a dataset with great initial camera poses, of a very reflective object. Just from the images, I can't get hloc to reconstruct the complete model. However, I have access to camera poses reconstructed by another software (metashape) without any keypoints, though. I've exported the poses to binary files only -no database-, which is what I've seen is exported by colmap_from_nvm too. I've checked that the coordinates makes sense, by plotting my results and partial hloc output on the image below.

As per your recommendation, I've tried running the Aachen pipeline, but since I don't have any keypoints, the [triangulator is failing with the message]: F0719 05:22:43.179829 17765 sfm.cc:473] Check failed: bundle_adjuster.Solve(&reconstruction)

I've tried adding some dummy arbitrary points, but it fails nonetheless: F0719 06:38:30.631529 7450 sfm.cc:473] Check failed: bundle_adjuster.Solve(&reconstruction)

For a subset of the images, I've plotted both the prior poses and the result of a partial reconstruction I got from hloc (after rotating the images according to their EXIF). Their forms seem to match pretty well, and therefore looks like the problem shouldn't be because of a faulty coordinate transform.

hloc poses are in red and metashape poses in white. Axis color coding are RGB for xyz.

How can I use these prior poses to be refined by hloc, especially in highly reflective objects on white background?

Thank you!

EDIT: Github's image posting was not working.