Clarification on the Implementation for Splatfacto Pruning Condition

[Tavish9]

hi, @jb-ye and @brentyi, thank you for PR #3376, which simplified Splatfacto using the gsplat strategy.

I have a question regarding the default strategy and the original implementation for pruning GSs using alpha contition, although some time has passed.

In the gsplat implementation, the pruning condition in gsplat/strategy/default.py is defined as follows:

self.reset_every = self.config.reset_alpha_every * self.config.refine_every
if step % self.reset_every == 0:
  reset_opa(
      params=params,
      optimizers=optimizers,
      state=state,
      value=self.prune_opa * 2.0,
  )

However, in the original implementation in splatfacto v1.1.4, the condition is:

reset_interval = self.config.reset_alpha_every * self.config.refine_every
if self.step < self.config.stop_split_at and self.step % reset_interval == self.config.refine_every:
    reset_value = self.config.cull_alpha_thresh * 2.0
    self.opacities.data = torch.clamp(
        self.opacities.data,
        max=torch.logit(torch.tensor(reset_value, device=self.device)).item(),
    )

It seems tha the resetting behavior has been moved slightly forward. Could you clarify the main considerations behind this change?

Furthermore, in official implementation of 3dgs, the condition is the same as what gsplat does:

if iteration % opt.opacity_reset_interval == 0 or (dataset.white_background and iteration == opt.densify_from_iter):
    gaussians.reset_opacity()

I am also curious to understand why the original implementation specifically uses self.config.refine_every in the condition.

I also conducted experiments using versions v1.1.4 and v1.1.5 with the same code and dataset, and I encountered something wired.

Although the final results were similar between the two versions, there were some notable differences:
1. In v1.1.5, the pruning of GSs was less aggressive, resulting in a significantly larger number of GSs being retained.
2. When I modified the alpha condition in v1.1.5 to match the condition used in v1.1.4, the behavior of the two versions became almost the same.

I would appreciate any insights you have into these differences and the potential reasons behind them.