search

fatPeter / mini-splatting

103 stars 10 forks source link

Question about Intersection Preserving #3

Closed yzhang-gh closed 4 months ago

yzhang-gh commented 4 months ago

Hi, I have tried this method and it is amazing!

However, I couldn't find the code related to the "Intersection Preserving" part. As far as I can see, there are two simplification iterations (i.e. simp_iteration1 and 2). Both of them make use of the imp_score of the gaussians (the former does downsampling and the latter cuts on a percentile). Am I missing something?

fatPeter commented 4 months ago

Thank you for your interest.

For simplicity, we merged Intersection Preserving into Importance-weighted Sampling, as:

imp_score[accum_area_max==0]=0

Here, accum_area_max means the 'maximum contribution area' of the corresponding Gaussian.

In the section of Intersection Preserving, we only preserve intersection points (i.e., Gaussians with accum_area_max!=0). In other words, we remove all Gaussians with accum_area_max==0. In this implementation, we simply set prob (imp_score) as 0, thus we can just ignore these Gaussians during sampling.

yzhang-gh commented 4 months ago

I see. Many thanks!