Image Data Augmentation

Introduction

This script is used to augment image data created using LabelMe-MIT.

It also creates new json files for the newly generated images, i.e., it augments both the image as well as it's annotation.

Constraints

LabelMe should be used for annotating images
Annotation should be a closed polygon/bounding box
There should be one annotation in an image

This script can work for images with multiple annotation as well, but it will only take into account the first annotation or the first user-specified class annotation.

Description

It copies the annotated portion from the reference image(input annotated image), processes it according to the user's instructions, which can be provided in the YAML file. And then paste it on one of the given random background images.

iseg_aug_yaml.py
input.yaml

Transforms

Downscale
Upscale
Rotation
Horizontal Flip
Vertical Flip
Random Shift
Blur - Averaging, Gaussian Blurring, Median Blurring, Bilateral Filtering
Noise - Gauss, Salt and Pepper, Poisson, Speckle
Grayscale
Brightness and Contrast
Canny Edge Detection

Options other than transforms

Threshold Ratio - Ratio of annotated area to background image area. Combinations below this ratio will be neglected.
User Class - Choose a specific class on which to do transformations
Pad Annotation - Amount of padding you want to add to the annotation