Noise Guided Splatting (NGS) Transparency Datasets

GitHub | Project Page

This repository contains the datasets used in the paper "Fix False Transparency by Noise Guided Splatting". It is designed to facilitate research and benchmarking for the "false transparency" artifact in 3D Gaussian Splatting (3DGS) reconstructions of opaque objects.

The repository is composed of four distinct subsets, each augmented with noise Gaussian infills (inside_gaussians.ply) crucial for evaluating surface opacity.

Dataset Description

The collection includes two original high-resolution datasets (stones and objects) and two augmented subsets from popular benchmarks (DTU and OmniObject3D). The primary purpose is to provide data exhibiting pronounced transparency issues and the necessary tools (noise infills) to quantify them using our proposed Surface Opacity Score (SOS) metric.

Subsets

1. Stones: A high-resolution object-centric dataset of over 100 stone specimens, captured with complex geometries and textures to challenge reconstruction robustness.
2. Objects: A supplementary dataset featuring a mixture of everyday objects with diverse material properties.
3. DTU: An augmented subset of the DTU Robot Image Data Set, with noise infills generated to evaluate transparency on these standard benchmarks.
4. OmniObject3D: An augmented subset of the OmniObject3D Dataset, similarly complemented with noise infills.

Dataset Structure

The dataset is organized into four main directories, one for each subset. Each scan within these directories follows a consistent structure:

.
├── stones/
│   ├── scan\_*/
│   │   ├── images/
│   │   ├── masks/
│   │   ├── sparse/0/
│   │   ├── inside_gaussians.ply        \# Noise Gaussians for evaluation
│   │   └── surface_gaussians.ply       \# Reconstructed surface Gaussians
│   └── ...
├── objects/
│   ├── scan\_*/
│   │   └── ...
├── DTU/
│   ├── scan\_*/
│   │   └── ...
└── OmniObject3D/
├── scan\_*/ (e.g., antique\_004, dinosaur\_004)
│   └── ...

Usage

This dataset is designed to be used with the Hugging Face datasets library, which can load each subset using a specific configuration name.

To evaluate transparency using our method, see the official NGS repository

Licensing

This dataset is released under a mixed license scheme:

• The stones and objects datasets are original works and are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
• The generated noise infill files (inside_gaussians.ply) for all subsets, including DTU and OmniObject3D, are also released under CC BY 4.0.
• The image and COLMAP data in the DTU and OmniObject3D subsets are provided here as derived works for convenience. They remain subject to their original licenses. Please consult the original dataset pages for specific licensing details.

Citation

If you use this dataset or the NGS methodology in your research, please cite our paper:

@inproceedings{ElHakie2025NGS,
    author    = {El Hakie, Aly and Lu, Yiren and Yin, Yu and Jenkins, Michael and Liu, Yehe},
    title     = {Fix False Transparency by Noise Guided Splatting},
    booktitle = {The Thirty-ninth Annual Conference on Neural Information Processing Systems},
    year      = {2025}
}