Dive into Illuminant Estimation from a Pure Color View

Abstract

Illuminant estimation is critically important in computational color constancy, which has attracted great attentions and motivated the development of various statistical- and learning-based methods. Past studies, however, seldom investigated the performance of the methods on pure color images (i.e., an image that is dominated by a single pure color), which are actually very common in daily life. In this paper, we develop a lightweight feature-based Deep Neural Network (DNN) model—Pure Color Constancy (PCC). The model uses four color features (i.e., chromaticity of the maximal, mean, the brightest, and darkest pixels) as the inputs and only contains less than 0.5k parameters. It only takes 0.25ms for processing an image and has good cross-sensor performance. The angular errors on three standard datasets are generally comparable to the state-of-the-art methods. More importantly, the model results in significantly smaller angular errors on the pure color images in PolyU Pure Color dataset, which was recently collected by us.