Domain adaptation based on optimal transport

Python package to simplify OT-based domain adaptation. This package uses POT as optimal transport backend.

Install with pip

$ python3 -m pip install otda --user

Install from source

$ python3 setup.py install --user

Exemplary code snippet

import otda

# Read source and target images
source_im = cv2.imread('source.jpg')
target_im = cv2.imread('target.jpg')

# Perform domain adaptation
adapted_im = otda.colour_transfer(source_im, target_im, method='linear', nsamples=1000)

Run domain adaptation on a single image

$ python3 -m otda.run --source source.jpg --target target.jpg --output output.jpg --method emd

Available methods: linear, gaussian, sinkhorn, emd.

Some examples of the colour adaptation

Source image	Target domain image	Method	Output	Computation time
		Linear		0.086s
		Gaussian (n=1000)		19.194s
		Sinkhorn (n=1000)		30.736s
		Earth mover's distance (n=1000)		22.988s

License

This repository is shared under an MIT license.

Author

Luis Carlos Garcia Peraza Herrera (luiscarlos.gph@gmail.com), 2020-2022.