A-WetDri: Adverse Weather Driving Dataset

Dataset URL: https://huggingface.co/datasets/unai-gurbindo/A-WetDri

---

📌 Overview

A-WetDri is a high-resolution, curated dataset developed to advance the robustness and generalization of perception models in autonomous driving under adverse weather conditions. It comprises both simulated scenes generated in CARLA and real-world driving footage from BDD100K and ACDC, incorporating challenging environmental scenarios such as:

• 🌧️ Rain
• 🌫️ Fog
• 🌃 Night
• ❄️ Snow
• 🌤️ Clear weather (baseline)

This resource is intended to support research in:

• Robust object detection under weather perturbations
• Domain adaptation between synthetic and real domains
• Weather-aware model training and evaluation
• Denoising and generative preprocessing methods for perception systems

---

🔧 Dataset Structure

The dataset is organized into two main frameworks, each comprising a comprehensive collection of images and metadata stored in .parquet format for portability and ease of access.

📦 Data Format

Each .parquet file contains the following fields:

{
  "file_name": "Town01_001.png",
  "image": image,
  "bbox": [[x_center, y_center, w, h], ...],
  "categories": [int, ...],
  "weather_condition": "fog"
}

📌 Note: Bounding boxes follow the YOLO format: [x_center, y_center, width, height], with all values normalized to the range $[0, 1]$ relative to image dimensions.

🧠 Object Classes

• Car
• Truck
• Bus
• Bicycle
• Motorcycle
• Walker
• TrafficSigns
• TrafficLight

🧩 Available Configurations

The dataset is divided into eight configurations, each corresponding to a specific framework (Simulated or Real), data type (annotations or images), and experimental setting (baseline or augmented). Below is a detailed breakdown:

🧪 Experimental Splits (Annotations + Metadata)

These configurations include object detection labels, weather metadata, and image dimensions. Useful for training and evaluating detection models.

• simulated_baseline: CARLA driving scenes under default weather conditions.
• simulated_augmented: CARLA driving scenes augmented with weather effects (fog, rain and night).
• real_baseline: real-world clear-weather scenes from BDD100K and ACDC.
• real_augmented: real-world scenes augmented using InstructPix2Pix diffusion model to simulate adverse weather (snow, fog, rain and night).

🖼️ Raw Images (Grouped by Weather)

These configurations contain only raw images, categorized by weather condition. Useful for generative tasks, domain adaptation, and visual inspection.

• simulated_trainable_images: synthetic training images by weather type.
• simulated_evaluation_images: synthetic evaluation images by weather type.
• real_trainable_images: real-world training images by weather type.
• real_evaluation_images: real-world evaluation images by weather type.

✅ How to Load a Specific Configuration

Use the load_dataset function from Hugging Face Datasets and specify the configuration with the name argument:

from datasets import load_dataset

# Example: Load the training split from the real augmented dataset
dataset = load_dataset(
    "unai-gurbindo/A-WetDri",
    name="real_augmented",
    split="train"
)

# Access image and bounding boxes
image = dataset[0]["image"]
bboxes = dataset[0]["objects"]["bbox"]

You can customize both the name and split depending on your task.

📋 Valid Configuration Combinations

name	split values	Description
simulated_baseline	train, val, test	Synthetic baseline experiment under clear conditions
simulated_augmented	train, val, test	Synthetic scenes with adverse weather augmentation
simulated_trainable_images	default, fog, night, rain	Raw synthetic images for training, by weather
simulated_evaluation_images	default, fog, night, rain	Raw synthetic images for evaluation, by weather
real_baseline	train, val, test	Real-world scenes with original weather (baseline)
real_augmented	train, val, test	Real-world scenes with simulated weather conditions
real_trainable_images	default, fog, night, rain, snow	Raw real-world images for training, by weather
real_evaluation_images	default, fog, night, rain, snow	Raw real-world images for evaluation, by weather

🖼️ Example: Visualizing Annotations from A-WetDri

This example shows how to load the dataset from Hugging Face, select a random annotated image from the simulated_augmented configuration, and draw bounding boxes around detected objects using the PIL library.

from datasets import load_dataset
import numpy as np
import random
from PIL import ImageDraw
import matplotlib.pyplot as plt

# Load the dataset configuration
dataset = load_dataset("unai-gurbindo/A-WetDri", "simulated_augmented")

# Select a random sample from the training split
random_index = random.randint(0, len(dataset["train"]) - 1)
sample = dataset["train"][random_index]

# Extract image and annotations
image = sample["image"]

bboxes = sample["bbox"]
labels = sample["categories"]

# Define class mapping
categories = ['Walker', 'Car', 'TrafficSigns', 'TrafficLight', 'Truck', 'Bicycle', 'Motorcycle', 'Bus']
id2label = {index: name for index, name in enumerate(categories)}
label2id = {name: index for index, name in id2label.items()}
color_map = {
    'Walker':      'cyan',
    'Car':         'lime',
    'TrafficSigns':'yellow',
    'TrafficLight':'magenta',
    'Truck':       'orange',
    'Bicycle':     'blue',
    'Motorcycle':  'white',
    'Bus':         'red'
}

# Draw bounding boxes on the image
draw = ImageDraw.Draw(image)
width, height = image.size

for i, box in enumerate(bboxes):
    x_center, y_center, w, h = box
    x_center *= width
    y_center *= height
    w *= width
    h *= height
    color = color_map.get(id2label[int(labels[i])], 'red')

    x_min = x_center - w / 2
    y_min = y_center - h / 2
    x_max = x_center + w / 2
    y_max = y_center + h / 2

    draw.rectangle([x_min, y_min, x_max, y_max], outline=color, width=1)

# Display the image
plt.figure(figsize=(10,10))
plt.imshow(image)
plt.axis('off')  # opcional: para ocultar los ejes
plt.show()

This script will display one randomly selected image from the simulated_augmented configuration, with bounding boxes of a colour based on the class labels of each detected object.

---

🔬 Evaluation & Augmentation Code

We are currently preparing the release of:

• 🧪 Evaluation scripts to assess object detection performance across weather conditions.
• 🌦️ Data augmentation tools for generating synthetic weather from clear-day samples using diffusion-based and image-to-image methods.

This functionality will be linked to this dataset card upon publication to ensure full reproducibility of results. (In progress)

---

📊 Dataset Statistics

Framework	Total Images	Train/Val	Test
Simulated	19,344	18,328	1,016
Real-World	23,046	22,156	1,250

Additional metrics (e.g., object frequency or Co-occurrence matrix) will be detailed in the accompanying publication.

🔄 Note: This table may be periodically updated as new images are added to the dataset or as part of routine quality control. In some cases, images may be removed due to labeling or generation errors. However, these adjustments will never affect the test sets, which are kept stable to ensure fair and consistent evaluation across experiments.

---

🧪 Reproducibility

• All training, validation, and test partitions are predefined and version-controlled.
• Annotations and augmentations are included for each sample.
• Compatible with datasets.load_dataset() from Hugging Face Datasets.
• Designed for immediate use in benchmarking pipelines and reproducible research workflows.

---

🤝 Contributing

We welcome contributions from the research community, including:

• Additional driving scenes under extreme weather
• Enhanced annotations or metadata corrections
• Integration of new augmentation strategies

To contribute, feel free to open an issue or submit a pull request via the dataset repository.