Deep Color Constancy Using Multi-Band NIR

This paper proposes deep a deep color constancy method that exploits multi-band NIR. In recent years, NIR col-orization has been popularly studied in literature. Our previous work shows that multi-band NIR is more useful than a single band NIR in the conversion of NIR-to-RGB. The proposed method exploits this relationship in illuminant estimation. The proposed network consists of two subnets that learn local illuminant and its confidence weight respectively. With multi-band NIR images, we can improve the confidence of local illuminant estimation. While most conventional methods focus on only spatial information of RGB image, we propose to utilize deep spatio-spectral features for color constancy. The experimental results show that the proposed multi-band NIR method performs better than the methods with RGB only or RGB + one band NIR.