Planckian Jitter data augmentation

Official implementation of the code from "Planckian Jitter: countering the color-crippling effects of color jitter on self-supervised training".

Now available in Kornia library!

Dependencies

Code written in Pytorch v1.8.1.

Pillow == 7.2.0
numpy == 1.19.5
torch == 1.8.1

Usage

Example usage with other torchvision transforms:

import torchvision.transforms as tranforms
from PIL import Image
from planckianTransforms import PlanckianJitter

img = Image.open('./demo_img/flw.jpg')

data_transforms = transforms.Compose([
transforms.RandomResizedCrop(size=img.size[0]),
transforms.RandomHorizontalFlip(p=0.5),
transforms.RandomVerticalFlip(p=0.5),
transforms.RandomApply([PlanckianJitter(mode="blackbody")], p=0.8)])

img_out = data_transforms(img)

img_out.show()

Run python3 ./demo.py to run that demo code.

Two parameters can be passed to the transform:

mode: choose between BlackBody points sampling or CIED. [Default is BlackBody]
idx: if idx is provided a specific illuminant is used instead of a random one from the sampled list.

# randomly selects an illuminant from the BlackBody list.
PlanckianJitter(mode="blackbody")

# randomly selects an illuminant from the CIED list.
PlanckianJitter(mode="cied")

# selects the illuminant at index 5 of the BlackBody list.
PlanckianJitter(mode="blackbody", idx=5)

Reference

If you are going to use this code please cite us:

@article{zini2022planckian,
         title = {Planckian jitter: enhancing the color quality of self - supervised visual representations},
         author = {Zini, Simone and Buzzelli, Marco and Twardowski, Bart{\l}omiej and van de Weijer, Joost},
         journal = {arXiv preprint arXiv: 2202.07993},
         year = {2022}
         }

License