On filtering gaussians by size

[leblond14u]

trafficstars

Hi GS community,

I'm trying to filter gaussians by size to get rid of the giant gaussians that tend to get in my scene during the training. To get a rough first try at solving my problem I tried this :

    def filter_GS_by_size(self):
        magnitudes = torch.norm(self._scaling, dim=1)
        threshold = torch.quantile(magnitudes, 0.95)
        valid_points_mask = magnitudes < threshold
        return valid_points_mask

And then adding this mask to the pruning algorithm in densify_and_prune() :

        ...
        self.valid_size_mask = self.filter_GS_by_size()
        prune_mask = torch.logical_or(prune_mask, ~self.valid_size_mask)
        self.prune_points(prune_mask)

However it is not working the way I was picturing it. Those big gaussians are not suppressed in the process and the parts of the scene that were well learnt before are now pruned as well in a weird fashion.

Has anyone succeeded in filtering the gaussians ?

Best,

Hugo