Color constancy from a pure color view

Shuwei Yue and *Minchen Wei

Color, Imaging, and Metaverse Research Center, The Hong Kong Polytechnic University.

If you use this code, please cite our paper:

@article{yue2023color,
  title={Color constancy from a pure color view},
  author={Yue, Shuwei and Wei, Minchen},
  journal={JOSA A},
  volume={40},
  number={3},
  pages={602--610},
  year={2023},
  publisher={Optica Publishing Group}
}

Code

Prerequisite

• Pytorch
• opencv-python

Training

• step 1: Preparing the dataset

To train PCC, training/validation data should have the following formatting:

 datasets/numpy_data/
	|1_8D5U5524.npy
	|2_8D5U5525.npy
	...
 datasets/numpy_labels/
 	|1_8D5U5524.npy
	|2_8D5U5525.npy
	...

---

Update on Sep. 29, 2024, There are some differences between the two formats:

• The Thumbnail version has raw data in png format, and the label is in json format, which includes extracted information from a 24-color card.
• The Full-resolution version has both raw and label in npy format, where the label contains only RGB information of the light source.

It is recommended to directly use the full-resolution version for training.

---

So, it is better to preprocess your data and corresponding labels into .npy format. The processed Recommend-ColorChecker dataset (CC2018) is provided in the folder of datasets/CC2018/. Noted that the black level and masked card of this dataset have been subtracted, then resized to $64\times64$ with normalized.

• step 2:

Run train.py

You can change the training fold number in config/param_config.py

Testing

move your trained model from log folder to the pretrain_models

Run test.py

PolyU Pure Color dataset

Thumbnails(.png) and full-resolution images(.npy) are provided in the following link:

PolyU PureColor dataset