search

w-m / 3dgs-compression-survey

An open survey on 3D Gaussian Splatting compression methods
https://w-m.github.io/3dgs-compression-survey/
MIT License
82 stars 4 forks source link

Add num Gaussians to table/data #5

Open w-m opened 3 months ago

w-m commented 3 months ago

Pruning and low-impact filtering are a large part of achieving high compression rates. Currently in the survey it's opaque how the different methods achieve their scores - by being able to represent the scene with very few Gaussians, or by achieving high compression of the attributes? I think we should start collecting the number of Gaussians per dataset. Will be interesting to compare in the table. And also enables us to build some graphs of PSNR over bits-per-Gaussian.

How to do it? @myla @paulbzm, should we just add a "#Gaussians" column to the .csv's? Just the number of (final) Gaussians, or is there some other interesting/related statistic?

If we decide on a format, I'll look up the data for Self-Organizing Gaussians, and we can add it to the data template. May need to ask other authors for the data (or decompressed scenes), I don't think most papers currently report the number.

w-m commented 3 months ago

Would also allow to look at PSNR/LPIPS/SSIM over number of Gaussians (no MB size) - to compare pruning and filtering methods, that don't necessarily do any compression.