vector-to-yolo

VecToYOLO: Geospatial Training Data Generator

VecToYOLO is a powerful QGIS plugin designed to streamline the creation of deep learning datasets. It automates the process of converting large-scale satellite imagery or drone orthomosaics into localized image “chips” (patches) with corresponding label files, formatted specifically for the YOLO (You Only Look Once) object detection and segmentation framework.

Core Features

• Dual Task Support: Supports both Object Detection (bounding boxes) and Instance Segmentation (polygon-to-mask conversion).
• Flexible Data Handling: * 8-bit Drone Orthomosaics: Standard RGB processing.
  • 16-bit Satellite Imagery: Automatic reflectance scaling (e.g., Sentinel-2 L2A) to 8-bit PNG for model compatibility.
• Advanced Data Splitting: Choose between Random splitting or Spatial splitting (using geographic rows) to ensure more robust validation and prevent data leakage.
• Customizable Overlap: Control the overlap percentage between chips—crucial for small object detection or continuous landcover mapping.
• Ready-to-Train Output: Generates a structured directory with train/ and val/ folders, a classes.txt file, and a pre-configured data.yaml file tailored for Kaggle, Google Colab, or Local environments.

How it Works

1. Alignment: Automatically reprojects your vector annotations to match the CRS of your raster.
2. Tiling: Iterates through your specified map extent, creating image tiles based on your desired chip size and overlap.
3. Label Mapping: Maps vector attributes to class IDs and converts geographic coordinates into the normalized YOLO format.