Large scale MasterSfM

[VladimirYugay]

Hey there,

Can you recommend script arguments or maybe a whole script to run the reconstruction of 1-5k images with Mast3r SfM?

In the demo, the amount of files is limited. Also, the reconstruction freezes after some time.

[abysee]

I encountered the same error, I just found that it's due to large system RAM occupation and the system would automatically kill the process. The easiest way is to reduce the size of input images but I'm wondering that there might be a way to train with smaller epoch or something. That way you'll have to modify the code rather than using the demo.

[kraidiky]

I have a similar question. I tried to build a model on 230 images taken with a phone camera, using the code from demo.py. running it directly, not through the demo interface. The model crashed with a message about insufficient cuda memory in the sparse_ga.py file line:
stacked_radius = (stacked_xy.view(B, 2, -1, H, W) - xy[:, :, None, S, S]).norm(dim=1)

I really want to see how to work with such or much more images. Inermentally or some other way.