3DGS Densification Config for SplaTAM

[Ramseyous0109]

Hi, thanks for this brilliant work.
I came across a bug running an offline demo on the replica dataset. I found that in the config file, use gaussian splatting-based densification is set to false:

use_gaussian_splatting_densification=False, # Use Gaussian Splatting-based Densification during Mapping

When I set this config to True, a bug appeared when training:

File "/home/user/splatam/SplaTAM/utils/slam_external.py", line 101, in accumulate_mean2d_gradient    variables['means2D_gradient_accum'][variables['seen']] += torch.norm(IndexError: The shape of the mask [853079] at index 0 does not match the shape of the indexed tensor [853064] at index 0

And this shape mismatch error can be reproduced in other dataset demos.

[pengchongH]

Same problem.

Thanks for your work. I also would like to know whether setting use_gaussian_splatting_densification=False means that we don't split the 3D gaussian.

[Nik-V9]

Hi, Thanks for your interest in our work!

We don't use the official 3DGS Densification in SplaTAM. We don't currently support the use_gaussian_splatting_densification config. It was used in an old version of our code. We will try to fix this bug.

[Ramseyous0109]

I get it. Thanks for your reply anyway!

[pengchongH]

Thanks for your reply!

[Buffyqsf]

Hello guys, I've meet a similar request. I haven't tried to run this densification strategy. But after reading the code, I don't know why this would happen. Have you fixed this bug?