In color constancy: data mattered more than network

Abstract

The objective of this paper is to argue that data mattered more than network in terms of color constancy. Computational color constancy is a linear operation device-dependent, which is part of the camera imaging pipeline. We extend the dataset based on this pipeline and prove that the scene illumination can be predicted using a very simple network as long as the dataset is large enough and evenly distributed. In the process of expanding the dataset, firstly, we remove illumination color casts in images which is ground-truth illumination color and then casts randomly generated evenly distributed illumination colors in images. We randomly generate five labels for each image and then work on the image to obtain this dataset. Using this dataset, we introduce a very simple network that is able to compute the color mapping function to correct the image’s colors. Experiments on our new datasets demonstrate that the method of this paper significantly outperforms the state-of-the-art color constancy methods.