Dataset by KRI-SOFT. He align it with RealityScan. I trained the rest with splatfacto-mcmc
Nerfstudio version 1.1.5 and Gsplat 1.5.2. Just 30k after about 60 minutes.
Eval images every 10th images. PSNR = 22.79, SSIM = 0.615
It has 3 Millions splats.
The result is really good
It can more good by removing eval dataset. So, All images will be trained instead of splits it into train-eval dataset.
The original PLY might better since this is compressed PLY with SOGS