CARLA-Seg

Synthetic Urban Scene Segmentation Dataset Generated with CARLA UE5

 

---

Table of Contents

1. Overview
2. Dataset Structure
3. Annotation Format
4. Statistics
5. License
6. Citation
7. Acknowledgements

---

Overview

CARLA‑Seg is a synthetic semantic‑segmentation dataset built with the latest CARLA UE5 simulator. It targets urban driving research scenarios where collecting highly‑annotated real data is costly or impractical. The dataset offers:

• 8 680 RGB images (7 232 train / 1 448 val) at 1280 × 720.
• Pixel‑perfect masks for 9 classes following a YOLOv11‑style segmentation format.
• Content generated from a 26‑hour continuous drive in the Upgraded Town 10 map with dynamic traffic (50 vehicles, 300 pedestrians) under daytime–sunny conditions.
• Realistic graphics via Lumen global illumination and Nanite geometry, narrowing the sim‑to‑real gap.

The goal is to provide a compact yet high‑quality benchmark for:

• Pre‑training and fine‑tuning semantic‑segmentation models.
• Synthetic‑to‑real domain‑adaptation studies.
• Rapid prototyping where real data is unavailable.

---

Dataset Structure

CARLA-Seg/
├── train/
│   ├── images/           # 7 232 RGB frames (.png)
│   └── labels/           # YOLOv11 masks (.txt or .json)
├── val/
│   ├── images/           # 1 448 RGB frames
│   └── labels/
├── carlos_coco.yaml      # Class names & palette
├── examples/             # Example RGB + mask pairs
├── scripts/              # Data‑generation & analysis scripts
└── docs/                 # Figures, banner, paper, etc.

Class List (carlos_coco.yaml)

ID	Name	Instances
0	car	Yes
1	motorcycle	Yes
2	bicycle	Yes
3	trafficlight	Yes
4	bus	Yes
5	person	Yes
6	roads	No
7	sidewalks	No
8	trafficsigns	Yes

---

Annotation Format

Masks follow YOLOv11 segmentation specs:

<class-index> <x1> <y1> <x2> <y2> ... <xn> <yn>.

---

Statistics

• Total objects: 364 186 (train + val).
• Avg objects / img: 42.
• Pixel coverage: roads 19.5 % > sidewalks ≈1.7 % > cars ≈1.5 %.
• Under‑represented class: bus (≈0.28 % of instances) – consider class‑balancing techniques.

---

License

The dataset and code are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license. You are free to share and adapt the material for any purpose, provided you give appropriate credit.

See LICENSE for details.

---

Citation

If you use CARLA‑Seg in your research, please cite the accompanying paper:

@misc{ruiz2025carlaseg,
  title  = {CARLA‑Seg: A Synthetic Urban Scene Segmentation Dataset Generated with CARLA UE5},
  author = {Ruiz, Carlos},
  year   = {2025},
  howpublished = {Dataset on GitHub},
  url    = {https://github.com/IronDog421/CARLA-Seg},
  note   = {Version 1.0}
}

---

Acknowledgements

• Developed with the open‑source CARLA Simulator.

Feel free to open issues or pull requests. Contributions are welcome!