muzafferardauslu/vehicle-3dgs

Vehicle-Class-Specific 3D Gaussian Splatting Dataset

🌐 Project Page  |  📄 Paper  |  🤗 Dataset

A collection of 114 vehicle 3D Gaussian Splatting (3DGS) models in .ply format, each corresponding to a distinct vehicle class (make / model / body type / year). The models were produced by an automated multi-stage reconstruction pipeline from heterogeneous, multi-source multi-view images. This is one of the first publicly available datasets of vehicle-class-specific 3D Gaussian Splatting reconstructions derived from in-the-wild image collections.

Dataset Summary

• Format: Gaussian Splatting .ply (binary) — one file per vehicle class
• Number of models: 114
• Modality: 3D Gaussian Splatting point sets (position, scale, rotation, opacity, spherical-harmonics color)
• Intended tasks: novel view synthesis, vehicle appearance modeling, damage assessment, 3D vehicle reconstruction benchmarking, generative AI and autonomous-systems research

How the models were built

The reconstruction pipeline has three stages:

1. Preprocessing — background removal and adaptive histogram equalization (CLAHE) to isolate vehicle regions and normalize illumination.
2. Orientation classification — an EfficientNet-B0 model sorts each image into one of eight directional views.
3. Sparse reconstruction — hierarchical localization + COLMAP with a neighbor-based image pairing strategy, followed by 3D Gaussian Splatting optimization.

See the paper for full methodology and evaluation.

Usage

Each .ply file is a standard 3D Gaussian Splatting file and can be opened with common 3DGS viewers and libraries (e.g. gsplat, SuperSplat, the original 3DGS renderer, or any viewer that reads Gaussian Splatting PLY).

from huggingface_hub import snapshot_download

path = snapshot_download(
    repo_id="muzafferardauslu/vehicle-3dgs",
    repo_type="dataset",
)
print(path)

To load a single splat file with plyfile:

from plyfile import PlyData

ply = PlyData.read("path/to/vehicle_001.ply")
print(ply.elements[0].data.dtype.names)

Directory structure

.
├── README.md
└── vehicles/
    ├── <vehicle_class_1>.ply
    ├── <vehicle_class_2>.ply
    └── ... (114 files)

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. You are free to share and adapt the material for non-commercial purposes, provided you give appropriate credit (cite both papers below).

Source data and usage terms

These 3D Gaussian Splatting models are derived from the CompCars dataset (MMLAB, The Chinese University of Hong Kong). The models were released publicly with the explicit permission of the CompCars authors. Because they are derived data, use of this dataset remains subject to the original CompCars terms, in particular:

• Non-commercial research use only. You may not sell, trade, or otherwise commercially exploit any portion of these models or data derived from them.
• Any publication using this dataset must also cite the original CompCars paper (see below).

CompCars original terms: https://mmlab.ie.cuhk.edu.hk/datasets/comp_cars/

Citation

If you use this dataset, please cite:

@article{uslu2026multistage,
  title   = {A Multi-Stage Pipeline for 3D Reconstruction of Vehicles from Heterogeneous Multi-View Images},
  author  = {Uslu, Muzaffer Arda and Kaplan Berkaya, Selcan},
  journal = {Erciyes {\"U}niversitesi Fen Bilimleri Enstit{\"u}s{\"u} Fen Bilimleri Dergisi},
  volume  = {42},
  number  = {2},
  year    = {2026},
  doi     = {10.65520/erciyesfen.1918173},
  url     = {https://doi.org/10.65520/erciyesfen.1918173}
}

Paper: https://doi.org/10.65520/erciyesfen.1918173

Acknowledgements

This work was supported by the Eskişehir Technical University Scientific Research Projects Commission under grant no. 25LÖT108.