Tipc & tricks for good model

graphdeco-inria / gaussian-splatting

Original reference implementation of "3D Gaussian Splatting for Real-Time Radiance Field Rendering"

https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

Other

12.47k stars 1.56k forks source link

Tipc & tricks for good model #825

Open smileyenot983 opened 1 month ago

smileyenot983 commented 1 month ago

Hello, everyone. Is there any set of tips & tricks on dataset acquisition for making visually appealing(without noise, blur etc) gaussian splatting?

For example, number of images for scene size, image resolution, maybe parameters of the optimizer

jaco001 commented 1 month ago

Sharp pictures...

smileyenot983 commented 1 month ago

Are there any tips on how to get rid of floating noise(attached example image) and to make everything more consistent? By consistency I mean if there are gaussians belonging to chair(or any other object), then it is usual case that when looking from some angles the surface of object is not smooth. I believe that depth supervision and more images could help. Is there anything else

jaco001 commented 1 month ago

There are a lot of things, but most of them are related to preparing a good dataset for photogrammetry. Colmap and GS take (literally) everything you give them, so weak images too. If you are thinking about taking photos indoors - this is a very good tutorial.

https://radiancefields.com/gaussian-splatting-brings-art-exhibitions-online-with-yulei https://medium.com/@heyulei/capture-images-for-gaussian-splatting-81d081bbc826

For me most important thing is sharp pictures with very deep depth of field