hierarchical mapper

[paul...@googlemail.com]

Hi,

I tried using the hierarchical mapper on some of my datasets. COLMAP seems to have problems with merging the models, the final output is multiple models in the specified output directory, which I also can't merge manually with the model merger. I tested the hierarchical mapper with 50 and 100 images overlap, 500 images per cluster. There is no output written to the console in the model merging step, so I'm not really sure what's happening. I run the hierarchical mapper with the same datasets (about 5000 images) as I would run the normal mapper. Feature matching with loop detection works nicely and the normal mapper produces really good models. I wanted to try the hierarchical mapper to avoid the 24h walltime limit on the HPC cluster that I'm running colmap on.

I know that the hierarchical mapper is not a final feature yet, but maybe you have some advise on how to use it? If it's not quite stable yet, maybe you have suggestions to reduce the time spend on bundle adjustment in the normal mapper? (Maybe reducing the number of iterations step by step? Would it help to fix the camera intrinsics with a shared camera after a few hundred images, to reduce the amount of parameters that have to be optimized?)

Thanks a lot!

[Johannes Schönberger]

Hi,

Did you inspect the individual models that could not be merged? Would you be able to share the dataset (with the database) for me to inspect?

You can speed up the bundle adjustment by limiting the iterations and by reducing the frequency at which global bundle adjustment runs (increase ba_global_images/points_ratio/freq). Otherwise, in case your intrinsic calibration is known, it helps to use the simplest model possible and to share intrinsics.

Cheers,

Johannes