not all poses recovered

[elliestath]

Hey, thanks for the great effort!

I am trying to evaluate my dense reconstruction pipeline.
However, for some datasets of the advanced set I cannot successfully recover all camera poses neither with COLMAP of OpenMVG.
I am actually missing 5 poses in the Palace set and 10 poses in the Auditorium.

My pipeline actually focuses on the dense reconstruction part, so I would need to start for some good poses, but I cannot recover them successfully with the available SfM pipelines.

Can I create the .log file anyway, so that I can upload and get the scores?

[arknapit]

I actually think you should be fine, the way the evaluation works (or at least how I remember it), is that it does a robust matching of camera poses, to figure out scale and rough alignment. And it does an ICP refinement on the dense PC afterwards, so if you have a reasonably good looking dense result, you should be fine. Did you try to submit your results already?

[arbab-isu]

@arknapit can you please point me to where in code it does the : "does a robust matching of camera poses, to figure out scale and rough alignment"?
grateful,

[arknapit]

it uses the Open3d registration function "registration_ransac_based_on_correspondence", which does not need the same amount of points (camera positions):

TanksAndTemples/python_toolbox/evaluation/registration.py

Line 65 in 3c2c212

def trajectory_alignment(map_file, traj_to_register, gt_traj_col, gt_trans,

you check the README if you want to find an example of it working: python_toolbox/evaluation/README.md