You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
When using RGB+D for rendering the depth, the depth in region or the image where there are not gaussians is zero.
This seems to be a somewhat arbitrary choice and in some cases we would like to use a positive constant value instead.
We can do that as a post-process using the rendered depth and the alpha value using rendered_depth + (1-alpha) * background_depth, but it would make things clearer if the gsplat rendering function had a background_depth argument that defaults to 0 and that would allow the user to provide a different background depth.
To implement this we would modify the line here
Can I take this up? The change seems simple enough and would be my first contribution.
However, I suggest using torch.full((C, 1), background_depth, device=backgrounds.device, dtype=backgrounds.dtype)
as that should remove the redundant filling with 1 and subsequent multiplication.
When using RGB+D for rendering the depth, the depth in region or the image where there are not gaussians is zero.
This seems to be a somewhat arbitrary choice and in some cases we would like to use a positive constant value instead.
We can do that as a post-process using the rendered depth and the alpha value using
rendered_depth + (1-alpha) * background_depth
, but it would make things clearer if the gsplat rendering function had abackground_depth
argument that defaults to 0 and that would allow the user to provide a different background depth.To implement this we would modify the line here
gsplat/gsplat/rendering.py
Line 719 in 7951619
by
[backgrounds, background_depth * torch.ones(C, 1, device=backgrounds.device)], dim=-1
The text was updated successfully, but these errors were encountered: