coco-transform-util

A python package to perform same transformation to coco-annotation as performed on the image.

Installation

Way 1

$ git clone https://github.com/Cargill-AI/coco-transform-util.git
$ cd coco-transform-util
$ python3 setup.py install

Way 2

$ pip3 install git+https://github.com/Cargill-AI/coco-transform-util.git
<<< Username: <user_id or email>
<<< Password: <personal access token or SSH key>

Personal Access token looks like this 83b318cg875a5g302e5fdaag74afc8ceb6a91a2e.

Reference: How to generate Personal Access token

Check installation

import ctu
print(ctu.__version__)

Benefits and Use Cases

1. Faster Model Training: Decrease the size of images and accordingly its annotation will be changed using this.
2. Flexibility: Rescaling of images and annotations to meet the need of Model/Framework.
3. Cost Saving: Lesser Computation requirement as images can be downscaled.
4. Interpretability: Annotation Visualization is also a part of this package.
5. Data Augmentation: <more practical in future>
6. Ability to handle other cases: Added Functionality such as cropping or padding of the annotation can help in multiple other cases such as:
   • cropping out each object image & annotation from an original image
   • cropping unnecessary area to zoom in on some particular area.
   • converting images to 1:1 aspect ratio by using padding and/or cropping.

How to use it?

Core

There are four core modules inside that helps in performing operations on COCO Annotation. These can imported as shown below:

from ctu import WholeCoco2SingleImgCoco, Coco2CocoRel, CocoRel2CocoSpecificSize, AggreagateCoco  

It's recommended that you have look at samples/example_core_modules.py to understand and explore how to use these.

Wrapper

Making use of wrappers can also come in handly to perform multiple operations in a much simpler and interpretable manner using the functions provided below:

from ctu import (
    sample_modif_step_di, get_modif_imag, get_modif_coco_annotation, 
    accept_and_process_modif_di, ImgTransform, Visualize
)

It's recommended that you have look at samples/example_highlevel_function.py to understand and explore how to use these.

Some sample data has also been provided with this package at example_data/* to explore these functionalities.

Demo / Sample

A sample HTML created from Jupyter-Notebook, contating some sample results has been added to the path samples/Demo-SampleOutput.html.

Version History

• v0.1: Core Modules: WholeCoco2SingleImgCoco, Coco2CocoRel, CocoRel2CocoSpecificSize. External Dependency on AMLEET package.
• v0.2: Removed the dependency on AMLEET package. Develop Core Module: AggreagateCoco. Addition of field "area" under "annotations" in coco.
• v0.3: Completed: Remove the out of frame coordinates in annotation. Update & add fields in "annotation" > "images". Ability to create transparent and general mask create_mask. In Development: Ability to export transformed image, mask and annotation per image wise and as a whole too.

Future

• Update the image fields in "images" key. (done)
• Crop out the annotation which are out-of-frame based on recent image shape. (done)
• Annotation Visualization + Mask creation can become a core feature to this library. (done)
• Rotate 90 degree left/right.
• Flip horizontally or vertically.
• COCO to other annotation format can also be a feature to this package.

Push to pypi

$ python3 setup.py sdist
$ twine upload dist/*