linkedai / flip Goto Github PK

Synthetic Image generation with Flip. Generate thousands of new 2D images from a small batch of objects and backgrounds.

License: GNU General Public License v3.0

Python 100.00%

synthetic-data generated-labels computer-vision object-detection deep-learning hacktoberfest

flip's Introduction

Flip

Synthetic Data generation with Flip! Generate thousands of new 2D images from a small batch of objects and backgrounds.

Installation

Install Flip using pip:

pip install flip-data

Dependencies

Flip requires:

Python (>= 3.7)
Opencv (>= 4.3.0)
Numpy (>= 1.19.1)

Quick Start (Example)

To try Flip library you can run examples/data_generator.py. You will need to add background images and objects to compose your new training dataset, then place them in the following directories:

BACKGROUNDS_PATTERN = "examples/data/backgrounds/*"
OBJECTS_PATTERN = "examples/data/objects/**/*"

The main workflow in Flip is to create transformers and then execute them as follows:

## Import Flip transformers
import flip.transformers as tr

OUT_DIR = "examples/result"

...

## Create Child transformers
transform_objects = [
        tr.data_augmentation.Rotate(mode='random'),
        tr.data_augmentation.Flip(mode='y'),
        tr.data_augmentation.RandomResize(
            mode='symmetric_w',
            relation='parent',
            w_percentage_min=0.2,
            w_percentage_max=0.5
        )
    ]

## Create main transformer
transform = tr.Compose([
    tr.ApplyToObjects(transform_objects),
    tr.domain_randomization.ObjectsRandomPosition(
        x_min=0, y_min=0.4, x_max=0.7, y_max=0.7, mode='percentage'
    ),
    tr.data_augmentation.Flip('x'),
    tr.domain_randomization.Draw(),
    tr.labeler.CreateBoundingBoxes(),
    tr.io.CreateJson(out_dir=OUT_DIR, name='img_generate.jpg'),
    tr.io.CreateJson(out_dir=OUT_DIR, name='json_generated.jpg')
])

## Execute transformations
el = tr.Element(image=..., objects=...)
[el] = transform(el)

Transformers

The main transformers are:

Transformer
Compose
ApplyToObjects
ApplyToBackground
ApplyToCreatedImage

By the way, all Transformers will be executed over objects of class Element and will return a new transformed Element.

Data Augmentation

Flip: Flip the Element in x or y axis.
RandomResize: Change the size of an Element randomly.
Rotate: Rotate Element randomly.
Color: Change color space or the element color.
Brightness: Changes the brightness in the image.
Contrast: Changes the contrast in the image.
Saturation: Changes the saturation in the image.
Noise: Add noise to the element image.
CutOut: Remove a section of the element in the desired area.
RandomCrop: Cut the image randomly.

Random Domain

Draw: Draw objects over background Element to merge them into a new image.
ObjectsRandomPosition: Set Random positions to objects over background Element.

Labeler

CreateBoundingBoxes: Draw bounding boxes around the objects contained by a background Element.
CreateMasks: Creates the segmentation mask for the objects contained in a background element.

IO

SaveImage: Save a .jpg File with the new generated image.
SaveMask: Save a .jpg File with the new generated mask.
Json: Save generated Labels as a Json.
Csv: Save generated Labels as a CSV.

Want to Contribute or have any doubts or feedback?

If you want extra info, email me at [email protected]

Report Issues

Please help us by reporting any issues you may have while using Flip.

License

Flip License

flip's People

Stargazers

Watchers

flip's Issues

Find what happend with histogram_matching_h in utils

In utils.py file in the methid overlay_transparent the reference to histogram_matching_h is lost.

Find what happend.

Move all the transfomers in core to separte files

Move all the Transformers find in the core.py file to separate Files, similar to what we do in data_augmentaion package.

Please consider using a permissive license?

Thanks for the free code!

I would like to use this to generate toy data for unit tests in a CICD setting for closed source software to replace. Butterflies would be more pleasant to look at than the colored blobs we currently generate :) This library's code wouldn't be shipped outside our organization, so I believe that's permitted copyright-wise, but the GPL still makes me nervous; I could see myself screwing up some setting in our deployment logic, or someone other than me might decide the test suite should run at a customer after all.

Assuming you still own all the code, any chance you'd consider switching to a permissive license (Apache, BSD, MIT, whatever)?

BACKGROUNDS_PATTERN = "examples/data/backgrounds/*"
OBJECTS_PATTERN = "examples/data/objects/**/*"