Unofficial implementation of Copy-Paste method: Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation.

This repo is implementation for semantic segmentation. You can use also in instance segmentation.

We do it at VOC-like dataset. If your dataset is coco-like, you need extract mask from coco first.

More details see blog.

1. Random Horizontal Flip
2. Large Scale Jittering
3. Copy-Paste

1. choice source image and main image;
2. get annotations from source image;
3. rescale source image and it annotations;
4. paste source image and annotations to main image and annotations;
5. merge main annotations and source annotations;

If use coco-like dataset, you need run get_coco_mask.py first:

usage: get_coco_mask.py [-h] [--input_dir INPUT_DIR] [--split SPLIT]

optional arguments:
  -h, --help            show this help message and exit
  --input_dir INPUT_DIR
                        coco dataset directory
  --split SPLIT         train2017 or val2017

And then, run copy_paste.py:

usage: copy_paste.py [-h] [--input_dir INPUT_DIR] [--output_dir OUTPUT_DIR]
                     [--lsj LSJ]

optional arguments:
  -h, --help            show this help message and exit
  --input_dir INPUT_DIR
                        input annotated directory
  --output_dir OUTPUT_DIR
                        output dataset directory
  --lsj LSJ             if use Large Scale Jittering

If your dataset is voc-like, just run copy_paste.py.

example of get_coco_mask.py:

python copy_paste.py --input_dir ../dataset/coco2017 --split train2017

example of copy_paste.py:

python copy_paste.py --input_dir ../dataset/VOCdevkit2012/VOC2012 --output_dir ../dataset/VOCdevkit2012/VOC2012_copy_paste --lsj True

main image + source image:

main anno + source anno:

visualization: