A Convolutional Framework for Color Constancy

Official implementation of:
Buzzelli, Marco, and Simone Bianco.
"A Convolutional Framework for Color Constancy."
IEEE Transactions on Neural Networks and Learning Systems (2024).

@article{buzzelli2024convolutional,
  title={A Convolutional Framework for Color Constancy},
  author={Buzzelli, Marco and Bianco, Simone},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2024},
  publisher={IEEE},
  doi={10.1109/TNNLS.2024.3454484}
}

Preparing the dataset

Experiments are performed on the ColorChecker dataset [Gehler et al. 2008] using the original threefold cross-validation split, but the “recommended” images and ground truth data by Hemrit et al. [2020].
More details on data preparation found in the ./Data/ folder of this repository

Training

See launcher_train.sh for example trainings on multiple test folds.

LCF-A

python cfcc.py --name-exp=exp_LCFA --image-size=200 --intermediate-blocks=0 --pool-size=1 --W-last=1 --W=3 --sigma=-1 --test-sets=1 --save-output --njet=1 --inner-size=9 --mink-norm=1.000001 --epochs=1000 --njet-protection1=prelu --nl=prelu --njet-protection2=posrelu --mink-protection-pow=oneabs --validation-transform=isotropic --data-augmentation

ECF-A

python cfcc.py --name-exp=exp_ECFA --image-size=200 --intermediate-blocks=1 --pool-size=1 --W-last=1 --W=3 --sigma=-1 --test-sets=1 --save-output --njet=1 --inner-size=9 --mink-norm=1.000001 --epochs=1000 --njet-protection1=prelu --nl=prelu --njet-protection2=posrelu --mink-protection-pow=oneabs --validation-transform=isotropic --data-augmentation

ECF-B

python cfcc.py --name-exp=exp_ECFB --image-size=200 --batch-size=8 --intermediate-blocks=1 --pool-size=3 --W-last=2 --W=5 --sigma=-1 --test-sets=1 --save-output --njet=1 --inner-size=27 --mink-norm=1.000001 --epochs=1000 --njet-protection1=prelu --nl=prelu --njet-protection2=posrelu --mink-protection-pow=oneabs --validation-transform=isotropic --data-augmentation

Running replica of Low-level framework by van de Weijer et al.

Replicating the behavior of baseline models from the framework:
Van De Weijer, Joost, Theo Gevers, and Arjan Gijsenij.
"Edge-based color constancy."
IEEE Transactions on image processing 16, no. 9 (2007): 2207-2214.

WP

python cfcc.py --mode='test' --epochs=0 --name-exp=expTMP --validation-transform=none --W=-1 --W-last=1 --njet=0 --intermediate-blocks=0 --mink-norm=-1 --sigma=0 --nl=none --njet-protection1=none --njet-protection2=none --mink-protection-pow=none

GW

python cfcc.py --mode='test' --epochs=0 --name-exp=expTMP --validation-transform=none --W=-1 --W-last=1 --njet=0 --intermediate-blocks=0 --mink-norm=1 --sigma=0 --nl=none --njet-protection1=none --njet-protection2=none --mink-protection-pow=none

SoG

python cfcc.py --mode='test' --epochs=0 --name-exp=expTMP --validation-transform=none --W=-1 --W-last=1 --njet=0 --intermediate-blocks=0 --mink-norm=5 --sigma=0 --nl=none --njet-protection1=none --njet-protection2=none --mink-protection-pow=none

GGW

python cfcc.py --mode='test' --epochs=0 --name-exp=expTMP --validation-transform=none --W=-1 --W-last=1 --njet=0 --intermediate-blocks=0 --mink-norm=5 --sigma=2 --nl=none --njet-protection1=none --njet-protection2=none --mink-protection-pow=none

GE1

python cfcc.py --mode='test' --epochs=0 --name-exp=expTMP --validation-transform=none --W=-1 --W-last=1 --njet=1 --intermediate-blocks=0 --mink-norm=5 --sigma=2 --nl=none --njet-protection1=none --njet-protection2=none --mink-protection-pow=none

GE2

python cfcc.py --mode='test' --epochs=0 --name-exp=expTMP --validation-transform=none --W=-1 --W-last=1 --njet=2 --intermediate-blocks=0 --mink-norm=5 --sigma=2 --nl=none --njet-protection1=none --njet-protection2=none --mink-protection-pow=none