jkdxbns/autonomous-driving-carla

CARLA Autonomous Driving Dataset

Custom datasets for autonomous driving in CARLA simulator

Created for CMPE 789 - Robot Perception at Rochester Institute of Technology

---

📊 Dataset Overview

This repository contains two custom-generated datasets from the CARLA 0.9.15 simulator for training autonomous driving perception models:

Dataset	Task	Images	Format	Size
YOLO Dataset	Object Detection	4,000	YOLOv8/v11	~1.2 GB
UFLD Dataset	Lane Detection	10,000	TuSimple-like	~3.4 GB

---

🚗 YOLO Object Detection Dataset

Description

Custom object detection dataset generated from CARLA Town01 with optimized graphics settings. Contains annotations for vehicles, pedestrians, traffic lights, and speed limit signs.

Classes

Class ID	Class Name	Total Annotations	Description
0	vehicle	2,797	Cars, trucks, vans
1	pedestrian	3,329	Walking pedestrians
2	traffic_light	409	Traffic signals
3	speed_limit	43	Speed limit signs

Total Annotations: 6,578

Split Distribution

┌─────────────────────────────────────────────────────────────────┐
│                    YOLO Dataset Splits                          │
├─────────────────────────────────────────────────────────────────┤
│  Train (70.3%)  ████████████████████████████░░░░░░░░  2,812     │
│  Val (19.8%)    ████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░    790     │
│  Test (9.9%)    ████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░    398     │
└─────────────────────────────────────────────────────────────────┘

Split	Images	Vehicle	Pedestrian	Traffic Light	Speed Limit
Train	2,812	1,949	2,252	301	33
Val	790	562	712	77	5
Test	398	288	365	31	5
Total	4,000	2,797	3,329	409	43

Class Distribution Chart

Class Distribution (Total Annotations)
═══════════════════════════════════════════════════════════════════

pedestrian    ████████████████████████████████████████████  3,329 (50.6%)
vehicle       ███████████████████████████████████░░░░░░░░░  2,797 (42.5%)
traffic_light █████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░    409 (6.2%)
speed_limit   █░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░     43 (0.7%)

═══════════════════════════════════════════════════════════════════

Format

Standard YOLO format with normalized bounding boxes:

<class_id> <x_center> <y_center> <width> <height>

Example:

0 0.499978 0.660904 0.117395 0.253719
1 0.726564 0.544706 0.078319 0.045380
2 0.545117 0.490625 0.004297 0.028472

Directory Structure

yolo_dataset/
├── train/
│   ├── images/          # 2,812 JPG images (1640×590)
│   └── labels/          # 2,812 TXT label files
├── val/
│   ├── images/          # 790 JPG images
│   └── labels/          # 790 TXT label files
├── test/
│   ├── images/          # 398 JPG images
│   └── labels/          # 398 TXT label files
├── dataset.yaml         # YOLO configuration file
└── classes.json         # Class ID mapping

---

🛣️ UFLD Lane Detection Dataset

Description

Lane detection dataset in TuSimple-like format, generated from CARLA Town01 for training Ultra-Fast Lane Detection (UFLD) models. Contains polyline annotations for left and right lane boundaries.

Split Distribution

┌─────────────────────────────────────────────────────────────────┐
│                    UFLD Dataset Splits                          │
├─────────────────────────────────────────────────────────────────┤
│  Train (70%)    ████████████████████████████░░░░░░░░  7,000     │
│  Val (20%)      ████████░░░░░░░░░░░░░░░░░░░░░░░░░░░░  2,000     │
│  Test (10%)     ████░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░  1,000     │
└─────────────────────────────────────────────────────────────────┘

Split	Images	Percentage
Train	7,000	70%
Val	2,000	20%
Test	1,000	10%
Total	10,000	100%

Annotation Format

Each image has:

1. Polyline annotations (.lines.txt): X,Y coordinate pairs for lane boundaries
2. Segmentation labels (.png): Pixel-wise lane masks
3. Ground truth list (train_gt.txt, val_gt.txt): Image paths with lane existence flags

Ground Truth Format:

<image_path> <label_path> <lane1_exist> <lane2_exist> <lane3_exist> <lane4_exist>

Example:

/images/train/000000.jpg /labels/train/000000.png 1 1 0 0

(Two lanes detected: left and right)

Directory Structure

ufld_dataset/
├── annotations/
│   ├── train/           # 7,000 .lines.txt files
│   ├── val/             # 2,000 .lines.txt files
│   └── test/            # 1,000 .lines.txt files
├── images_train.zip     # 7,000 JPG images (2.3 GB)
├── images_val.zip       # 2,000 JPG images (647 MB)
├── images_test.zip      # 1,000 JPG images (325 MB)
├── labels_train.zip     # 7,000 PNG segmentation masks (18 MB)
├── labels_val.zip       # 2,000 PNG segmentation masks (5.1 MB)
├── labels_test.zip      # 1,000 PNG segmentation masks (2.6 MB)
└── list/
    ├── train_gt.txt     # Training split ground truth
    ├── val_gt.txt       # Validation split ground truth
    └── test.txt         # Test image list

Note: Images and labels are provided as ZIP files to reduce file count. Extract after downloading.

---

📥 Download Instructions

Using Hugging Face CLI

# Install huggingface_hub if needed
pip install huggingface_hub

# Download YOLO dataset (ready to use)
huggingface-cli download jkdxbns/autonomous-driving-carla yolo_dataset --repo-type dataset --local-dir ./

# Download UFLD dataset
huggingface-cli download jkdxbns/autonomous-driving-carla ufld_dataset --repo-type dataset --local-dir ./

Extract UFLD ZIP Files

cd ufld_dataset

# Extract images
unzip images_train.zip -d .
unzip images_val.zip -d .
unzip images_test.zip -d .

# Extract labels
unzip labels_train.zip -d .
unzip labels_val.zip -d .
unzip labels_test.zip -d .

Using Python

from huggingface_hub import snapshot_download

# Download entire dataset
snapshot_download(
    repo_id="jkdxbns/autonomous-driving-carla",
    repo_type="dataset",
    local_dir="./datasets"
)

---

🖼️ Image Specifications

Property	Value
Resolution	1640 × 590 pixels
Format	JPEG
Color	RGB
FOV	150° (wide-angle)
Camera Position	Front-mounted, 2.4m height
Simulator	CARLA 0.9.15
Map	Town01 (optimized graphics)

---

🎯 Intended Use

These datasets are designed for:

• Training object detection models (YOLO, Faster R-CNN, etc.) for autonomous driving
• Training lane detection models (UFLD, LaneNet, etc.)
• Research in simulation-to-real transfer learning
• Educational purposes in robotics and computer vision courses
• Benchmarking perception algorithms in controlled environments

---

📈 Training Results

Models trained on these datasets achieve:

Model	Task	Performance
YOLO11n	Object Detection	mAP50: 0.85+
UFLD ResNet-18	Lane Detection	Accuracy: 92%+

Pre-trained weights available at: 🤗 jkdxbns/autonomous-driving-carla

---

📚 Citation

If you use this dataset in your research, please cite:

@misc{carla_autonomous_driving_dataset_2024,
  author = {Jatin Khokhani},
  title = {CARLA Autonomous Driving Dataset},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/jkdxbns/autonomous-driving-carla}
}

---

📄 License

This dataset is released under the MIT License.

---

🔗 Related Resources

• GitHub Repository: autonomous-driving-carla
• Pre-trained Models: Hugging Face Models
• CARLA Simulator: carla.org
• UFLD Paper: Ultra Fast Structure-aware Deep Lane Detection

---

Created with ❤️ for CMPE 789 - Robot Perception @ RIT