Multiple Illumination Estimation with End-to-End Network

Most popular color constancy algorithms assume that the light source obeys a uniform distribution across the scene. However, in the real world, the illuminations can vary a lot according to their spatial distribution. To overcome this problem, in this paper, we adopt a method based on a full end-to-end deep neural model to directly learn a mapping from the original image to the corresponding well-colored image. With this formulation, the network is able to determine pixel-wise illumination and produce a final visually compelling image. The training and evaluation of the network were performed on a standard dataset of two-dominant-illuminants. In this dataset, this approach achieves state-of-the-art performance. Besides, the main architecture of the network simply consists of a stack of fully convolutional blocks which can take the input of arbitrary size and produce correspondingly-sized output with effective learning. The experimental result shows that our customized loss function can help to reach a better performance than simply using MSE.