data augmentation tool to generate data for object detection and segmentation (random synthesis)

Generated Images

Requirement

• python 2.7
• opencv-python
• numpy
• imutils

Data Organization

• data-augmentation
  • bg (background images)
    • 1.jpg
    • 2.jpg
    • ...
    • 100.jpg
  • dataset (save path)
    • Images: Synthesized rgb images
    • labels: labels for YOLO training
    • Masks: Masks for each object (segmentation data, maximum value=1)
  • images (object images without background)
    • 01_0001.png
    • 01_0002.png
    • ...
    • 02_0003.png

Note

• Object images and background images must be the same as given format.
• Object images must be cropped to fit the object.

Demo

• Run synthesis.py

Usage

1. Change config.json according to your configuration.

• NUM_MIN_OBJ: Minimum number of objects to be synthesized into one picture
• NUM_MAX_OBJ: Maximum number of objects to be synthesized into one picture
• SCALE_MIN: Minimum size of objects (pixel)
• SCALE_MIN: Maximum size of objects (pixel)
• NUM_IMAGES: Number of images to create

2. Run synthesis.py