Memory errors

[rwijmashell]

Hi, thank you for your great work!
I'm currently working on testing this on larger datasets (5-10k images) and notice that a very large amount of (V)RAM would be required to make it work. I've already generated a pairs file to reduce the number of pairs from 50M to 3M, but this still seems to be way too large. Do you have any pointers/suggestions I could try out to make it scale better?

I'm using cloud compute with 80GB VRAM and 220GB RAM so that shouldn't be an issue btw.

[davidblom603]

I have this problem as well. @yocabon any suggestions?

[yocabon]

Hi, the current global optimization strategy was not meant to be run on large collections of images. We are working on a better solution, but no guarantee at this point.

[zcczhang]

Hi, the current global optimization strategy was not meant to be run on large collections of images. We are working on a better solution, but no guarantee at this point.

@yocabon Do you have any suggestions for parameters suitable for relatively dense images (e.g., more than 100 images per scene)? I've tried different prefilter and scene_graph settings to reduce the total number of pairs to roughly ~1k, but none seem to work as effectively as when subsampling 32 images with a complete scene graph. Maybe longer alignment iterations or other? Thank you!