Applying inverse sigmoid for opacities in export_ply

[pschroeppel]

Hi,

in the original 3DGS repo, when saving Gaussians to a .ply file, the opacity is saved in its unactivated form (i.e. before applying a sigmoid function): https://github.com/graphdeco-inria/gaussian-splatting/blob/54c035f7834b564019656c3e3fcc3646292f727d/scene/gaussian_model.py#L246

Based on this, I would assume, that any 3DGS viewer that loads a .ply file, will apply a sigmoid to the loaded opacities. Hence it would make sense to apply an inverse sigmoid function ( https://github.com/graphdeco-inria/gaussian-splatting/blob/54c035f7834b564019656c3e3fcc3646292f727d/utils/general_utils.py#L18 ) to the opacities before storing them in https://github.com/dcharatan/pixelsplat/blob/59d420a409cd0f059366808a9f956001536b55f1/src/model/ply_export.py#L26 .

A small disclaimer: I didn't look into code of 3DGS viewers, to confirm that they apply a sigmoid on the loaded opacities. However, I've played a bit with NoPoSplat and there I saw clear artifacts when visualizing my exported .ply files, which disappeared when I applied the inverse sigmoid before exporting. When visualizing exported .ply files from pixelSplat, the improvement is a bit less obvious, but in my opinion it also looks better, with the inverse sigmoid:

• Without applying inverse sigmoid on the opacities before exporting (visualized using https://antimatter15.com/splat/):

• With applying inverse sigmoid on the opacities before exporting:

It would be great if you could confirm whether the inverse sigmoid indeed makes sense here and potentially update the code. Thanks and best, Philipp