vggt icon indicating copy to clipboard operation
vggt copied to clipboard
verified publisher icon facebookresearch

Comparison of 3DGS Training Results: Output from demo_colmap.py performs worse than COLMAP's output in training

Open zhp2018 opened this issue 5 months ago • 4 comments

hi, When using the COLMAP results generated by demo_colmap.py for ‌Gaussian Splatting (3DGS)‌ reconstruction, the reconstruction quality is worse than when using data directly converted from original COLMAP. However, the initial point cloud generated by demo_colmap.py is of better quality and higher density. This discrepancy is confusing. Mainly regarding the clarity of rendered images demo_colmap output Training Results:[ITER 10000] Evaluating train: L1 0.016932867839932444 PSNR 27.74833793640137 colmap output Training Results:[ITER 10000] Evaluating train: L1 0.012061026878654957 PSNR 32.750404357910156

zhp2018 avatar Aug 06 '25 02:08 zhp2018

Hi three options may matter here:

  1. Use bundle adjustment or not
  2. The confidence threshold you use
  3. How images are loaded

jytime avatar Aug 06 '25 13:08 jytime

Hi three options may matter here:

  1. Use bundle adjustment or not
  2. The confidence threshold you use
  3. How images are loaded

Bundle adjustment was not used Default parameters were used I'm not quite sure what you mean. I have 32 images, and the local path is set via scene_dir= Thank you!

zhp2018 avatar Aug 07 '25 09:08 zhp2018

https://github.com/facebookresearch/vggt/blob/8492456ce358ee9a4fe3274e36d73106b640fb5c/demo_colmap.py#L46

jytime avatar Aug 12 '25 23:08 jytime

Doesn't 3dgs training require undistorted images, and what we get here is undistorted with demo_colmap directly?

kuaiqushangzixiba avatar Sep 10 '25 10:09 kuaiqushangzixiba