Dataset

We introduce a new dataset of 503,875 RGB images, annotated according to 13 tasks (criteria).
Each task has its own set of classes, for a total of 56 classes.

The data are released under a CC-BY license.
Any use of this dataset should cite the associated paper (coming soon) or the authors
(see the license for more details).

• 📦 Images: (to be completed)
• 📝 Full annotations: (to be completed)
• 📝 Train annotations: (to be completed)
• 📝 Test annotations: (to be completed)

For a detailed description of how the train and test sets are constructed, see:
➡️ train-test-construction.md

---

Fig. 1 – Example of annotation. Each image is annotated on 13 criteria.

As illustrated in Fig. 1, each image is manually annotated by
three independent annotators according to 13 criteria, of which 12 are related to weather and 1 to the viewpoint (an informative criterion not used for training, intended to diversify the dataset in terms of camera perspectives).

Image resolutions range from 640×450 to 1280×720.
Images are extracted from public videos released under CC0 or Open Licence.

---

Annotation criteria (13 tasks)

• Weather Type
  Clear, Sunny+Clear, Rain, Snow, Fog, Fog+Rain, Fog+Snow, None

• Weather Intensity
  Low, Medium, High, None

• Visibility
  Very Low, Low, Medium, Good

• Sky Condition
  Unknown, Clear Sky, Partly Cloudy, Cloudy, Overcast, Partly Overcast

• Precipitation Presence
  None, Rain, Snow, Hail

• Precipitation Intensity
  None, Low, Medium, High

• Ground Condition
  Dry, Wet, Partly Wet, Snowy, Partly Snowy, Wet+Snowy, Unknown

• Glare / Reflections
  Absent, Present

• Light Conditions
  Day, Night, Sunset, Sunrise, Artificial Light

• Water Spray (road spray)
  Absent, Present

• Water on Windshield
  Absent, Present, None

• Snow on Windshield
  Absent, Present, None

• Viewpoint (not used for training)
  Onboard Vehicle, Pedestrian, Fixed Road Camera

---

Fig. 2 – Example images showing rain, snow, fog, and clear weather, by day and by night.

Fig. 2 shows a few examples of images from the dataset,
including rainy, snowy, foggy, and clear scenes, both during day and night.

---

Global dataset statistics

Fig. 3 – Overview of class distributions across all tasks.

Fig. 3 provides an overview of class distributions for all tasks.
For some tasks (e.g., presence of road spray or water/snow on the windshield), the data distribution is highly imbalanced.
We recommend taking this imbalance into account during training (e.g., class weighting, focal loss, etc.).

The following figures focus more specifically on Weather Type and Ground Condition.

---

Fig. 4 – Data distribution for the “Weather Type” task.

Fig. 4 shows the class distribution for the Weather Type task.
The distribution is globally well balanced, which is favourable for training robust classifiers on this task.

---

Fig. 5 – Data distribution for the “Ground Condition” task.

Fig. 5 shows the class distribution for the Ground Condition task.
Here as well, the distribution is reasonably balanced, which enables effective training of models for this specific task.

---

Annotation tools and figure generation

The Annotations.py script allows you to:

• perform your own annotations,
• and automatically generate the statistical figures shown above (global distributions, per-task distributions, etc.).

Examples of usage and how to generate these statistics will be added soon in the repository documentation.

1. Install dependencies

   In a terminal, at the root of the project:

   pip install -r requirements.txt
   

2. Watch the video

Annoted_images_small.mp4