What is the best way to create augmentation on image dataset while training instance segmentation?

Question

I am using YOLACT instance segmentation model to train set of images. The dataset size is very small (~20 images). The model doesn't converge properly (of-course given the dataset size). I wanted to increase the dataset size by adding some augmented images. I know we have various image augmentation techniques and packages like imgaug , albumentation, opencv etc. but I need image & annotation file (COCO JSON ) format to train the model.

Is there a package that helps me to automatically generate the annotations of augmented images?

or is there a better way to address my issue ?

Answer 1

Simple Copy Paste is a strong method for data augmentation for instance segmentation related tasks.

Check about the research paper here.

For unofficial github code, check here

Albumentation and TorMentor are also useful libraries for data augmentation.