albertsuarez/object-cut

✂️ Cut the main object of an image automagically

Website | RapidAPI | Demo | Status page | Paper

✂️ Cut the main object of an image automagically

This repository contains all the logic necessary to run the ObjectCut API.

Summary

Object Cut is an online image background removal service that uses BASNet. Removing the background from an image is a common operation in the daily work of professional photographers and image editors. This process is usually a repeatable and manual task that requires a lot of effort and human time. However, thanks to BASNet, one of the most robust and fastest performance deep learning models in image segmentation, Object Cut was able to turn it into an easy and automatic process.

It was built as an API to make it as easy as possible to integrate. APIs, also known as Application Programming Interfaces, are already a common way to integrate different types of solutions to improve systems without actually knowing what is happening inside. Specifically, RESTful APIs are a standard in the Software Engineering field for designing and specifying APIs. Making it substantially easier to adapt your desired APIs to your workflows.

Object Cut was born to power up the designing and image editing process from the people who work with images daily. Integrating the Object Cut API removes the necessity of understanding the complex inner workings behind it and automates the process of removing the background from images in a matter of seconds.

Requirements

1. Python 3.7+
2. Docker CE 19+
3. Docker-compose 1.27+

Recommendations

Usage of virtualenv is recommended for package library / runtime isolation.

Usage

To run the server, please execute the following from the root directory:

1. Set up environment creating the .env file. This file must have this structure (without the brackets):

   SECRET_ACCESS={SECRET_ACCESS}
   

2. Set up your Google Cloud Storage credentials, decrypt it using GPG with the needed passphrase and decompress it:

   gpg --quiet --batch --yes --decrypt --passphrase="{{ GPG_PASSPHRASE }}" --output ./multiplexer/keys/storage_key.tar ./multiplexer/keys/storage_key.tar.gpg
   tar xvf ./multiplexer/keys/storage_key.tar -C ./multiplexer/keys

3. Build everything in parallel:

   docker-compose build --parallel 

4. Deploy the whole stuck (multiplexer, inference and traefik) with just this command.

   docker-compose up -d --scale multiplexer=1 --scale inference=3

That's it! You have ObjectCut running on port 80 routing traffic using traefik.

Run tests

1. Run ObjectCut locally

2. Move to multiplexer module

   cd multiplexer

3. Run tests

   SECRET_ACCESS={SECRET_ACCESS} python3 -m unittest discover -v

Development

Change underlying model

This project was built using BASNet as the model for inferring the Salient Object Detection. However, in order to test other ones we added the support to select also the different versions of U^2-Net (U2NET, U2NETP and U2NETPORTRAIT), also implemented by Xuebin Qin, in the Inference container specifying it as a environment variable called MODEL. You can do that setting your model name at docker-compose.yml:

inference:
    image: object_cut_inference
    build: inference
    env_file:
      - .env
    expose:
      - 80
    volumes:
      - '/tmp:/tmp'
      - '/var/run/docker.sock:/var/run/docker.sock:ro'
    networks:
      - object_cut
    restart: always
    environment:
      - MODEL=BASNet  # Can also be `U2NET`, `U2NETP` or `U2NETPORTRAIT`

Integrations

This API integrates with several external APIs which are listed below.

Google Cloud Storage

For being able to upload the image response to a public bucket for let the users download the output, we are using GCS for doing it once the users specify that in the output_format=url (default value) in the API request.

The integration is pretty simple. Every request is being authenticated using the service account JSON file under the multiplexer/keys folder where every output image is being upload to the object-cut-images bucket under a Life Cycle policy of 3 days (this has been configured on the Google Cloud Console UI). Once the file has been uploaded (and make it public) the library itself it returns you the public URL that you can return to the user.

How to add a new test

Create a new python file on the multiplexer module called test_*.py in test.api.* with the following structure:

from test.base import BaseTestClass


class NewTest(BaseTestClass):

    def test_v0(self):
        expected = 5
        result = 2 + 3
        self.assertEqual(expected, result)

Authors

• Adrià Cabeza
• Albert Suàrez

License

Apache-2.0 © ObjectCut

object-cut's Issues

bug: Broken UI on website

Describe the bug
Entirely broken UI for the website page.

To Reproduce
Visit (https://objectcut.com/) on a desktop computer.

Expected behavior
Working UI

Screenshots
Imgur album for screenshots of broken UI

Additional context
Seems like a dependency issue that requires correction/rebuilding.

Can share a bit more about the post-processing?

Thanks for your great work!

can share some ideas about the post-processing module?
or does it also included in this repo?

Really appreciate if you can reply.

Have a nice day.

Selection of inference network

This is a very good project, may I ask why you didn't choose U2-net as the inference network of the project?

bug: Rotation based on meta-data

Describe the bug
Images with rotation meta-data should be rotated since currently every viewer tool display images in the correct rotation, which then ObjectCut is outputting the processed image rotated, which confuses the end-user.

To Reproduce
Based on the maracuja request on the RapidAPI discussion thread, sending an image meta-data rotated is outputting a result image with a different rotation.

Expected behavior
ObjectCut should return the image in the correct rotation, if possible.

Screenshots
This can be found in the RapidAPI thread specified above.

Additional context
None.