Non-Orthogonal Projection Method for Multiple Illuminant Estimation

We present a method to estimate the color of multiple illuminants in real-world images. The proposed method exploits a physical model for the color formation of a pixel and infers the probability of a pixel chromaticity with a specific color temperature on Planckian locus. A pair of pixels in its neighborhood in the 3-dimensional uniform color space votes for the probability, and the most probable direction of projection is found statistically. The proposed method does not require a pre-processing step such as color segmentation to obtain a uniform surface reflectance. The newly created multiple-illuminant image dataset demonstrated that the proposed method outperforms an orthogonal method in both the scenes consisting of dominant colored- and multiple colored-objects.