Skip to content

The SODA Dataset is a computer vision dataset containing aerial imagery of small objects captured at different altitudes. The dataset contains 829 images and 6719 object annotations.

Notifications You must be signed in to change notification settings

danielopisani/SODA-Dataset

Repository files navigation

SODA-Dataset

The SODA Dataset is a computer vision dataset containing aerial imagery of small objects captured at different altitudes. The Small Objects from Different Altitude (SODA) Dataset contains 829 images and 6719 object annotations. The objects are annotated using polygons.

Test

Download SODA Dataset

The SODA dataset can be downloaded from here.

  • Description: The 829 images are grouped into their respective altitude directory.
  • Recommended usage: Test and validation using images captured from different altitudes.
  • Preprocessing: No image pre-processing steps were applied to the images.
  • Augmentation: No image augmentation steps were applied to the images.
  • Description: The 829 images are merged into one directory.
  • Recommended usage: Training a small object detector using images captured at different altitudes.
  • Preprocessing: No image pre-processing steps were applied to the images.
  • Augmentation: No image augmentation steps were applied to the images.

Access the SODA Dataset from Roboflow

The SODA dataset is also hosted on Roboflow for public access. The SODA Data repository on Roboflow can be accessed from here.

Dataset Classes

  • Clear Plastic Bottles (clear_plastic_bottle)
  • Drink Cans (drink_can)
  • Drink Cartons (drink_carton)
  • Glass Bottles (glass_bottle)
  • Glass Jars (glass_jar)
  • Other Plastic Bottles (other_plastic_bottle)

SODA Dataset Composition

  • 1m (Captured using an iPhone SE): 452 images | 720 annotations
  • 5m: 202 images | 1853 annotations
  • 10m: 50 images | 1142 annotations
  • 15m: 35 images | 958 annotations
  • 20m: 30 images | 768 annotations
  • 25m: 30 images | 716 annotations
  • 30m: 30 images | 612 annotations

Altitudes are all AGL (Above Ground Level). Aerial images were captured using a combination of DJI Mini 2 and a DJI Air 2S drones.

SODA YOLOv8 Model

Image Preprocessing

  • Auto-Orient
  • Resize (640x40)
  • Tile (5x5)
  • Dropped 'unlabelled litter' class.

Image Augmentation

  • No image augmentation steps applied for model training.

News

January 27, 2024: First upload of SODA dataset with 829 images and 6719 annotations.

About

The SODA Dataset is a computer vision dataset containing aerial imagery of small objects captured at different altitudes. The dataset contains 829 images and 6719 object annotations.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published