metadata

annotations_creators:
  - expert-annotated
language:
  - en
language_creators:
  - found
license:
  - cc-by-nc-3.0
multilinguality:
  - monolingual
pretty_name: DAWN Dataset
size_categories:
  - 1K<n<10K
source_datasets:
  - original
tags:
  - object-detection
  - computer-vision
  - autonomous-driving
  - adverse-weather
  - traffic
task_categories:
  - object-detection
task_ids:
  - object-detection

Dawn Dataset

This dataset contains images with annotations for object detection in the YOLO format, converted to absolute coordinates for easier use. The dataset is focused on detecting vehicles and people in various environments.

Dataset Source

This dataset is based on the DAWN (Detection in Adverse Weather Nature) dataset, which was originally published by Mourad KENK on Mendeley Data:

Original Source: DAWN Dataset on Mendeley

The original DAWN dataset consists of real-world images collected under various adverse weather conditions (fog, snow, rain, and sandstorms). It emphasizes diverse traffic environments (urban, highway, and freeway) and provides a rich variety of traffic flow scenarios specifically for autonomous driving and video surveillance applications.

Dataset Structure

The dataset is split into training and validation sets and contains annotations for the following 6 classes:

Truck (class_id: 0)
Person (class_id: 1)
Bicycle (class_id: 2)
Car (class_id: 3)
Motorcycle (class_id: 4)
Bus (class_id: 5)

Dataset Format

Each example in the dataset includes:

image: The actual image data (PIL.Image format when loaded)
width: Original image width in pixels
height: Original image height in pixels
objects: A list of detected objects in the image with the following properties for each:
- class: The numerical class ID (0-5)
- class_name: The text name of the class (Truck, Person, etc.)
- x_min: The x-coordinate of the top-left corner of the bounding box (in pixels)
- y_min: The y-coordinate of the top-left corner of the bounding box (in pixels)
- width: The width of the bounding box (in pixels)
- height: The height of the bounding box (in pixels)

Coordinate System

Note that the original YOLO format uses normalized coordinates (relative to image size) with center points:

x_center, y_center: Normalized center point (0-1)
width, height: Normalized dimensions (0-1)

This dataset converts these to absolute pixel coordinates with top-left origin:

x_min, y_min: Top-left corner in pixels
width, height: Dimensions in pixels