alexmkwizu/gaussian_training_datasets

Gaussian Training Datasets (COLMAP) for msplat

COLMAP-format multi-view scenes for training 3D Gaussian Splatting models, packaged for msplat — a Metal-native 3DGS trainer for Apple Silicon. Also includes pre-trained .ply splats under tested_outputs/.

All scenes are redistributed from third-party datasets. Full credit goes to their original authors — see Licensing & credits and please cite the original papers. This repo only repackages them in COLMAP layout for convenience.

Contents

mipnerf360/{bicycle,bonsai,counter,garden,kitchen,room,stump}/   # Mip-NeRF 360
tandt/{train,truck}/                                             # Tanks & Temples
db/{drjohnson,playroom}/                                         # Deep Blending
    └── images/  +  sparse/0/{cameras,images,points3D}.bin       # COLMAP layout

tested_outputs/                # pre-trained 3DGS .ply splats (+ SUMMARY.md, RESULTS.md)

Usage with msplat

pip install -U "huggingface_hub[cli]"

# Download everything into ./datasets/
hf download alexmkwizu/gaussian_training_datasets --repo-type dataset --local-dir datasets

# Or a single scene
hf download alexmkwizu/gaussian_training_datasets --repo-type dataset \
    --include "tandt/truck/*" --local-dir datasets

# Train (pick -d by native image size: Mip-NeRF 360 ~16 MP -> -d 4; T&T/DB ~1 MP -> -d 1)
msplat datasets/mipnerf360/garden -n 7000 -d 4 --eval
msplat datasets/tandt/truck      -n 7000 -d 1 --eval

Pre-trained splats (tested_outputs/)

Standard 3DGS binary PLYs trained with msplat (7000 iters) on an M4 / 16 GB MacBook Pro. Indoor scenes reach PSNR 27–30. Drag any .ply into a web viewer such as SuperSplat to view. See tested_outputs/SUMMARY.md.

Licensing & credits

This dataset redistributes scenes from the following works. Each retains the license/terms of its original source — consult the original project pages, and if you use these scenes, cite the original papers.

Mip-NeRF 360 — mipnerf360/

Scenes from the Mip-NeRF 360 dataset (Google Research). Project page & terms: https://jonbarron.info/mipnerf360/

@inproceedings{barron2022mipnerf360,
  title     = {Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields},
  author    = {Barron, Jonathan T. and Mildenhall, Ben and Verbin, Dor and
               Srinivasan, Pratul P. and Hedman, Peter},
  booktitle = {CVPR},
  year      = {2022}
}

Tanks and Temples — tandt/ (train, truck)

From the Tanks and Temples benchmark (Intel). COLMAP-preprocessed version as distributed by Inria GRAPHDECO. Project: https://www.tanksandtemples.org/

@article{Knapitsch2017,
  title   = {Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction},
  author  = {Knapitsch, Arno and Park, Jaesik and Zhou, Qian-Yi and Koltun, Vladlen},
  journal = {ACM Transactions on Graphics},
  volume  = {36}, number = {4}, year = {2017}
}

Deep Blending — db/ (drjohnson, playroom)

From Deep Blending for Free-Viewpoint Image-Based Rendering (UCL / Inria). COLMAP-preprocessed version as distributed by Inria GRAPHDECO.

@article{hedman2018deep,
  title   = {Deep Blending for Free-Viewpoint Image-Based Rendering},
  author  = {Hedman, Peter and Philip, Julien and Price, True and Frahm, Jan-Michael
             and Drettakis, George and Brostow, Gabriel},
  journal = {ACM Transactions on Graphics (SIGGRAPH Asia)},
  volume  = {37}, number = {6}, year = {2018}
}

COLMAP preprocessing (Tanks & Temples + Deep Blending)

The COLMAP versions of the Tanks & Temples and Deep Blending scenes are those distributed with the 3D Gaussian Splatting project, Inria GRAPHDECO: https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/

@article{kerbl3Dgaussians,
  title   = {3D Gaussian Splatting for Real-Time Radiance Field Rendering},
  author  = {Kerbl, Bernhard and Kopanas, Georgios and Leimk{\"u}hler, Thomas and
             Drettakis, George},
  journal = {ACM Transactions on Graphics}, volume = {42}, number = {4}, year = {2023}
}

COLMAP (Structure-from-Motion)

Camera poses / sparse points were produced with COLMAP (Schönberger & Frahm, CVPR 2016; Schönberger et al., ECCV 2016): https://colmap.github.io/

---

Trained-splat outputs in tested_outputs/ were generated by msplat (Apache-2.0). The input scenes remain under their original licenses as above.