low quality for MipNeRF-360 dataset

[ZERO-black]

Hi, thank you for your great work!

I tried the garden scene from the MipNeRF-360 dataset, but the result with full training was not very good.
As shown in the video (video1), the background is quite blurry, and there are noticeable long Gaussian artifacts.

I’ve read this issue and tried training a single chunk using train_single.py, which gave a much better result (video2).
So I suspect that the problem comes from train_coarse.py, which generates the initial scaffold and skybox.
Do you know why this step might degrade the background quality? I'd appreciate it if you could share any ideas or hypotheses.

Also, even when training chunks only (without train_coarse.py), I noticed that long Gaussians appear much more frequently in H3DGS compared to 3DGS. This was also mentioned in the issue above, but the discussion hasn’t been updated for a while.

Is this behavior expected in H3DGS, or could it be related to the training commands or viewer I used?
If this is indeed a limitation of H3DGS, do you have any insights into why these long Gaussians appear more often?

Here are the commands I used:

Full training: (video1)

python scripts/full_train.py --project_dir ${DATASET_DIR}

Single chunk training: (video2)

python -u train_single.py -s ./data/mipnerf360/garden/camera_calibration/chunks/0_0 --model_path data/mipnerf360/garden/output/chunks/0_0 -i ../../rectified/images -d ../../rectified/depths --skip_scale_big_gauss    


submodules/gaussianhierarchy/build/GaussianHierarchyCreator data/mipnerf360/garden/output/chunks/0_0/point_cloud/iteration_30000/point_cloud.ply ./data/mipnerf360/garden/camera_calibration/chunks/0_0  data/mipnerf360/garden/output/chunks/0_0


python train_post.py -s data/mipnerf360/garden/camera_calibration/chunks/0_0 --model_path data/mipnerf360/garden/output/chunks/0_0 --hierarchy data/mipnerf360/garden/output/chunks/0_0/hierarchy.hier --iterations 15000 --feature_lr 0.0005 --opacity_lr 0.01 --scaling_lr 0.001

For rendering, I used SIBR_viewers/install/bin/SIBR_gaussianHierarchyViewer_app for both case.

Thanks in advance!

[ameuleman]

Hi,
Yes, it looks like the scaffold has issues dealing with small scenes. I am not sure about the elongated Gaussians; depth loss might hurt results in scenes with a lot of coverage (controlled with --depth_l1_weight_init).