Hacker News Reader

Currently, what's the best way to make gaussian splat scenes from video?

Thats quite a difficult question, given the large amount of methods (https://github.com/MrNeRF/awesome-3D-gaussian-splatting).

I think "3D Gaussian Splatting as Markov Chain Monte Carlo" looks very promising, as it redesigns the training so that it is no longer dependent on the desification algorithm: https://arxiv.org/abs/2404.09591v1

Ah, I'm rather looking for a ready-to-go repo for a layman like me to just have a quick go. The latest I used was nerfstudio a long time ago. Seems to be updated regularly still, I wonder if it supports training splat scenes now. I also wonder if there were any better similar projects since then.

If your video source is of a static scene that the camera is moving through. You could try poly.cam

Upload you video directly to the website for Gaussian splat processing

And based on their license acknowledgments poly.cam uses nerfstudio and gsplat, in case anyone else was wondering

https://learn.poly.cam/third-party-libraries

Nerfstudio dev here! We have a splatfacto method in nerfstudio that provides a similar interface to our nerf methods.

That's good - I'm tired of the racial bias that seems to come out of a lot of ML work these days - any algorithm which makes everyone Indian just isn't going to work for most users

Reply on news.ycombinator.com