Skip to content

Latest commit

 

History

History
40 lines (31 loc) · 1.78 KB

README.md

File metadata and controls

40 lines (31 loc) · 1.78 KB

Copy & Paste Data Augmentation

Original dataset images should be saved into <Constants.DATA_DIR>/.

Synthetic images will be saved into <Constants.DATA_DIR>/synthesized/.

You will need a csv file listing all background images (see /data/zero_lid.csv for reference).

You will need a csv file listing all images with exactly one object (see /data/one_lid.csv for reference). Please note that our method does not work with very dark objects/low contrast between background and object.

Start Copy & Paste Data Augmentation by creating DataAugmentation object, with the paths to the two csv files, in /data_augmentation/.

from data_augmentation.data_aufmentation import DataAugmentation
data_augmentation_object = DataAugmentaion(zero_lid_dir, one_lid_dir)

Then run generate method on this DataAugmentation object. You can switch on different tags:

  • noise_background: add noiso to the images with label 0 (switched on by default)
  • rotate: rotate some inserted objects with random angle (switched on by default)
  • change_color: change color of some inserted objects
  • make_edge: allows creating object in the edges
  • make_dark: make some inserted objects dark
  • make_transparent: make some inserted objects transparent
  • coco_annotation: create a coco annotation json file with annotations for instance segmentation

For example the following code generates a new synthetic data set with 6 classes, 10 images in each class and all tags switched on:

dataAugm.generate(
    classes={0: 10, 1: 10, 2: 10, 3: 10, 4: 10, 5: 10}, 
    noise_background=True, 
    rotate=True, 
    change_color=True, 
    make_edge=True, 
    make_dark=True, 
    make_transparent=True, 
    coco_annotation=True,  
    data_dir_name='train'
    )