How to work on large scale scenes? about sfm-disambiguation-colmap

[lxxue]

Thanks for you interest. Unfortunately, the largest dataset I have worked on is berliner_dom with 1.6 k images only. So I am not sure if the code can run with 20k images without some problems like OOM. My suggestion would be try this Kataria's method first (you can use our modification to compile it with latest versions of COLMAP) since it is basically a COLMAP with some modifications. COLMAP should scale well and the runtime of this method should be close to a usual COLMAP run as it doesn't introduce many computations. Still, I think it will take days to run a scene at this scale and tuning parameters for it would be painful.

In the meantime, you can also try the Cui's method as I implement a parallel version of it on CPU. I am not sure how long this prefiltering will take but it might be good for the colmap mapper because it remove quite some matches beforehand. I am not sure if its runtime (prefiltering + COLMAP reconstruction) will be faster than a vanilla COLMAP run.

One friend also said that SuperGlue is helpful for repetitive scenes but I haven't try that much.

[zhangmozhe]

Thanks for your detailed feedback so much! Thanks for your patience and great work!