Illuminant Equivariant Networks for Computational Color Constancy

Neural networks are now the standard solution to many computer vision problems. Their generalization ability enables them to successfully address various tasks in computational photography, such as enhancement, restoration, and color constancy. However, their performance is highly dependent on the illumination conditions of the training images. When faced with test images under different illuminant conditions, these networks often struggle to perform their tasks correctly. In this paper, we investigate the efficacy of illuminant equivariant neural networks for the illuminant estimation task, which is crucial for computational color constancy. These networks are equivariant to the photometric transformations that characterize changes in lighting conditions. They achieve this capability through mathematical derivation rather than specific augmentation during training. We implemented the equivariant versions of state-of-the-art neural networks for illuminant estimation and tested them on the NUS dataset. The results demonstrate that the equivariant networks maintain stable performance even with significant changes in illumination, whereas the original standard networks exhibit a serious degradation in their accuracy.