MIST

Multi-Domain Synthetic Dataset for Rural Driving

Jongwon Ryu, Jaehoon Go, Trung X. Pham, and Junyeong Kim

MIST is a synthetic rural-driving dataset for studying environmental domain variation. This repository accompanies the paper and provides documentation and dataset-access examples. The image data is hosted on Hugging Face, not in GitHub.

This repository was previously named SMS and is now named MIST to match the published paper and dataset.

Overview

Unlike urban-centric driving data, MIST focuses on rural roads and variations in background appearance. Scenes are generated with Slowroads and organized into 32 balanced domain configurations:

Factor	Values
Season	Spring, summer, autumn, winter
Time of day	Dawn, daytime, dusk, night
Weather	Clear, overcast

The dataset supports research on multi-domain image-to-image translation, vision-language analysis, and environmental domain shifts. Domain labels are provided as text, such as autumn dawn clear weather rural road.

Public Data

The current Hugging Face release provides image-text pairs in Parquet format:

Split	Image-text pairs
Train	32,000
Test	3,200
Total	35,200

Field	Content
image	Rural-driving image, decoded as a PIL image when loaded
text	Text description of the environmental domain

The released files occupy approximately 73 GB compressed. Streaming is recommended for inspecting a few samples without downloading the full release. These counts describe the current public subset, not the full dataset described by the paper.

The dataset card lists segmentation annotations and remaining data as ongoing release work. The current default configuration contains only image and text; do not assume that segmentation masks are already included. Check the dataset card and files for the latest availability.

Quick Start

Use Python 3.10 or newer:

git clone https://github.com/jongwonryu/MIST.git
cd MIST
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# Save three streamed test images and their domain descriptions.
python examples/preview_dataset.py --split test --limit 3 --output outputs/preview

The requirements pin the tested Datasets version to avoid a reported partial-Parquet-stream shutdown issue in newer scanner releases.

Or use Hugging Face Datasets directly:

from datasets import load_dataset

dataset = load_dataset(
    "jongwonryu/MIST-autonomous-driving-dataset",
    split="test",
    streaming=True,
)
sample = next(iter(dataset))
print(sample["text"])
print(sample["image"].size)
sample["image"].save("mist_sample.png")

Image dimensions should be read from the loaded sample rather than assumed from the simulator's original rendering settings. Set streaming=False only when you intend to download and cache the requested split locally.

Repository Scope

This GitHub release contains dataset documentation, citation metadata, and a small preview utility. It does not currently contain the simulator modifications, dataset-generation pipeline, or training/evaluation code used in the paper.

Citation

@article{ryu2026mist,
  title   = {{MIST}: Multi-Domain Synthetic Dataset for Rural Driving},
  author  = {Ryu, Jongwon and Go, Jaehoon and Pham, Trung X. and Kim, Junyeong},
  journal = {IEEE Access},
  volume  = {14},
  pages   = {132866--132877},
  year    = {2026},
  doi     = {10.1109/ACCESS.2026.3725755}
}

License and Contact

The Hugging Face dataset card declares Apache-2.0 for the dataset. No separate software license has been specified for this GitHub repository. Refer to the dataset card and applicable simulator terms when reusing those materials.

Jongwon Ryu: fbwhddnjs511@cau.ac.kr.