search

OceanYing / OmniSeg3D-GS

3D Gaussian Splatting adapted version of OmniSeg3D (CVPR2024)
MIT License
55 stars 2 forks source link

Setting of parameters for Object Loss #1

Closed dhgras closed 6 months ago

dhgras commented 7 months ago

Hi, thank you for your excellent work!

I am trying to apply this work to my own dataset, which contains over 1000 rgb images with a resolution of around 2k. I have some doubts about the object loss:

  1. When does the average_flag need to be set to True?
  2. How to adjust clustersize/batchsize?

Looking forward to your reply!

OceanYing commented 7 months ago

Hi, thanks for your interest in our work!

  1. With average_flag = False, the sampled pixel number in each patch $p$ is proportional to the pixel-size of $p$. Since this sampling strategy may ignore the small patches, you can turn on the average_flag to sample the same number of pixels in each patch. This is a simple trial and we found average_flag = False performs better in most cases empirically. Therefore I leave the default setting as average_flag = False.
  2. Larger clustersize / batchsize usually results in faster convergence since more pixels are taken into contrastive learning in each iteration. You can set the value depending on your GPU memory.
  3. For your case, I would recommend the default setting (average_flag = False and a large value of batchsize >=32768). We have tested on LERF and MIP360 datasets with this setting and it should work for you, too.

Feel free to contact me if you have further questions!

dhgras commented 7 months ago

Okay, I got it. Thank you for your reply!