Undesirable results when using large datasets

YihangChen-ee / HAC

:house: [ECCV 2024] Pytorch implementation of 'HAC: Hash-grid Assisted Context for 3D Gaussian Splatting Compression'

Other

197 stars 11 forks source link

Undesirable results when using large datasets #8

Closed jxbb233 closed 3 months ago

jxbb233 commented 4 months ago

Hi, I tried to train using the kitti360 dataset (>3000 images) and the results don't seem to be very good as shown below. I tried manually adjusting the voxel size and iteration, but not many changes. I feel the problem is that there are not enough Gaussians in the scene to fit the image. Is there any other parameter I can try to adjust? Thanks for your help!

YihangChen-ee commented 3 months ago

Hi, Thanks for being interested in our work! I am afraid there are not enough anchors/Gaussians in this scene. As HAC is built upon Scaffold-GS, have you tried running this dataset using Scaffold-GS?

jxbb233 commented 3 months ago

It turns out that the problem was caused by an error in our data preprocessing with COLMAP. After switching methods, the issue no longer occurs.