Implementation of anisotropic regularization on gaussians

[haoyuhsu]

Hi,

I am really impressed by this work and still figuring out the details. I would like know if you have kept the implementation of anisotropic regularization in the codebase or not. Thanks a lot!

[a1600012888]

I added this part to the original gaussian splatting training code.

# add anisotropic reg:
gaussian_scaling = gaussians.get_scaling # [N, 3]
aniso = torch.max(gaussian_scaling, dim=1)[0] / (torch.min(gaussian_scaling, dim=1)[0] + 1e-5)
aniso_clip = torch.clamp(aniso, args.aniso_thres, None) - args.aniso_thres
loss_aniso = torch.mean(aniso_clip)

[haoyuhsu]

Thanks for your sharing!