Supporting code to the paper
O Sidorov. Artificial Color Constancy via GoogLeNet with Angular Loss Function
The code is designed for MATLAB R2017b with Neural Networks Toolbox. It should also work with other recent versions.
Running the code requires having installed pre-trained GoogLeNet model. (Can be installed within MatLab using Add-On explorer, or externally from File Exchange)
Most of the operations required for modification the network, loading the data, and performing a training, is contained in main.m.
angularRegressionLayerL#.m contains description of regression layers with custom loss function. Which one to use can be specified in main.m, line 7 (L2 by default).
This file contains original illumination vectors real_illum_11346.mat normalized by their length (so the new length is unit).