search

nerfstudio-project / gsplat

CUDA accelerated rasterization of gaussian splatting
https://docs.gsplat.studio/
Apache License 2.0
2.16k stars 273 forks source link

Gsplat rendering has holes #330

Open smart4654154 opened 2 months ago

smart4654154 commented 2 months ago

Thank you for your work. I am using a viewer（ https://github.com/hangg7/nerfview ） Looking at the PLY data, I found some holes. image

Specifically, My PLY data is generated by 3DGS, https://github.com/graphdeco-inria/gaussian-splatting . The Nerfview viewer used gsplat to render data and i found some holes. I guess it may be due to the gsplat algorithm:

  1. I used the built-in viewer SIBR in 3DGS to render.Everying is good. My training data is matrix city (big city) https://city-super.github.io/matrixcity/ image

  2. I used the CityGS viewer(https://github.com/DekuLiuTesla/CityGaussian) , and it displayed normally. Please note that their rendering package is diff-Gaussian-rasterization from 3DGS. Then I was using the code of citygs to read PLY data, and rendering the data using the gsplat, but there were still holes, so I guess it may be a problem with gsplatform

could you help me thank you

liruilong940607 commented 2 months ago

It might be the background color that is tricking your eyes. From the images you post it seems like you were using black background to render when using SIBR viewer but white background when using gsplat. Have you tried rendering with gsplat but set background to black?

It might be the case that your model is just transparent on some area but looks ok with a black background.

smart4654154 commented 2 months ago

It might be the background color that is tricking your eyes. From the images you post it seems like you were using black background to render when using SIBR viewer but white background when using gsplat. Have you tried rendering with gsplat but set background to black?

It might be the case that your model is just transparent on some area but looks ok with a black background.

That's right, you're right. Thank you very much By the way, may I ask I use nerfstudio's splatfacto algorithm to train the data(matrix city -big city). I use the nerfview tool（ https://github.com/hangg7/nerfview) ,which uses gsplatform , to render the data. an error message will appear: image image image Note: When an error occurs, the CUDA of the task manager suddenly reaches 100% I encountered an error on the 30th second of using nefview

I guess it's because the background color defaults to random in nerfstudio's splatfacto algorithm, And in the nerfview tool, backgrounds=torch,ones (1, 3, device=device), so this error is generated. What's your opinion?