Creating Synthetic Image Datasets

This tool helps create synthetic data for object detection modeling. Given a folder of background images and object images, this tool iterates through each background and superimposes objects within the frame in random locations, automatically annotating as it goes. The tool also resizes the icons to help the model generalize better to the real world.

Setup

Clone this repo. Then create and activate the conda environment provided:

$ conda env create -f environment.yml
$ conda activate images

Place background images in the Backgrounds/ subfolder and objects in the Objects/ subfolder.

Create

Run the create.py script to generate hundreds/thousands of synthetic training images for object detection models.

$ python create.py

Output images will be placed in the TrainingData/ subfolder once done.

Args

These are the available entrypoint arguments that you can supply at runtime. More will be added in the future.

• --backgrounds: Path to folder of background images.
• --objects : Path to folder of object images.
• --output : Path to folder of output images.
• --groups : Whether or not to place groups of objects together.
• --annotate : Whether or not to create and save annotations for the new images.
• --sframe : Whether or not to create a Turi Create SFrame for modeling.
• --mutate : Perform mutatuons to objects (rotation, brightness, shapness, contrast)