MUGSQA (Multi-Uncertainty-Based Gaussian Splatting Quality Assessment) Dataset

Officially released MUGSQA dataset for the ICASSP 2026 paper MUGSQA: Novel Multi-Uncertainty-Based Gaussian Splatting Quality Assessment Method, Dataset, and Benchmarks.

📃Dataset Summary

MUGSQA is a large-scale dataset designed for Gaussian Splatting Quality Assessment (GSQA). It is constructed by introducing multiple uncertainties during the reconstruction process and collecting large-scale subjective quality scores.

The dataset contains 2,414 reconstructed Gaussian models, each paired with rendered videos and Mean Opinion Scores (MOS). It supports research on:

• Gaussian Splatting quality assessment
• reconstruction robustness evaluation
• rendering-based and rendering-free quality metrics

The dataset simulates several uncertainties that commonly occur during reconstruction, including:

• input view resolution
• number of input views
• view-to-object distance
• initialization of the point cloud
• method of GS reconstruction

These factors create diverse reconstruction distortions that are useful for benchmarking reconstruction methods and quality metrics.

📁Dataset Structure

The dataset repository contains the following files:

reference/
main.tar.gz
additional.tar.gz
mos_main.xlsx
mos_additional.xlsx

1. Reference Videos

reference/

This folder contains ground-truth reference videos rendered from the original source objects. Due to copyright restrictions, the original 3D mesh models are not included. Only the rendered reference videos are provided. These videos are used in the subjective quality assessment experiments as reference stimuli.

2. Main Set

main.tar.gz

After extraction:

main/
├── sample_folder_1/
├── sample_folder_2/
...
└── sample_folder_1970/

The main set contains 1,970 reconstructed Gaussian objects. Each sample folder represents one distorted reconstruction generated under specific uncertainty settings.

2.1. Naming Convention

Each sample folder follows the format:

modelname_resolution_views_distance_method_pointcloud

Example:

12th-c-ce-water-moon-guanyin_480res_9views_5distance_lgs_rndpc

Where:

Field	Description
modelname	name of the source object
resolution	input view resolution
views	number of input views
distance	view-to-object distance
method	reconstruction method
pointcloud	initial point cloud type

These parameters correspond to the reconstruction uncertainty settings used during dataset generation.

2.2. Files inside each sample folder

Each distorted sample folder contains two files with the same name:

sample_name.mp4
sample_name.ply

File	Description
.mp4	rendered video of the reconstructed Gaussian object
.ply	reconstructed 3D Gaussian model

3. Additional Set

additional.tar.gz

After extraction:

additional/
├── sample_folder_1/
...
└── sample_folder_444/

The additional set contains 444 reconstructed samples. Unlike the main set, the additional set includes reconstructions generated using multiple Gaussian Splatting methods:

• 3DGS
• Mip-Splatting
• Scaffold-GS
• EAGLES
• Octree-GS

All other settings are consistent with the main set.

4. MOS Annotations

The subjective quality scores are stored in:

mos_main.xlsx
mos_additional.xlsx

Each entry corresponds to one distorted sample and its Mean Opinion Score (MOS). MOS values represent perceptual quality collected through a large-scale subjective study. Higher MOS indicates better perceived quality, and the MOS range is 0 to 5.

🧰Usage Example

Example workflow:

1. Extract the dataset archives:

tar -xzf main.tar.gz
tar -xzf additional.tar.gz

2. Load reconstructed models:

sample_folder/
├── sample_name.mp4
└── sample_name.ply

3. Use the .ply files for rendering-free quality assessment or use the .mp4 files for rendering-based quality assessment.

4. Match the sample name with the corresponding MOS score in the Excel files.

🔗Citation

@inproceedings{chen2026mugsqa,
  title={{MUGSQA: Novel Multi-Uncertainty-Based Gaussian Splatting Quality Assessment Method, Dataset, and Benchmarks}}, 
  author={Tianang Chen and Jian Jin and Shilv Cai and Zhuangzi Li and Weisi Lin},
  booktitle={Proc. ICASSP},
  year={2026}
}