Object-Color Description Under Varying Illumination

Given a fixed and uniform illumination, metameric objects appear as the same color. However, when the illumination is altered, two metameric reflecting objects under the first illuminant may no longer produce the same color signal under the second. This situation is called metamer mismatching. Metamer mismatching poses several challenges for the camera and display industries as well as color-based computer vision technology.In light of metamer mismatching, the present study criticizes the conventional approaches to color description when the illuminant alters, and then lays a foundation to robustly describe object colors under varying illumination conditions. Later, the degree of metamer mismatching is used as a measure of the quality of lights. We demonstrate that although the common color spaces such as CIELAB and related spaces in the literature may work well for a fixed illuminant, they can lead to poor results when the illuminant is changed. In view of these problems, new descriptors for hue, lightness and chroma are presented that are based on properties of a Gaussian-like spectrum metameric to the given color tristimulus coordinates. Experiments show that the new Gaussian-based appearance descriptors correlate with different descriptors as well as the CIECAM02 appearance model does on average. Furthermore, the Gaussian-based descriptors are significantly more stable than the descriptors defined in the CIECAM02 appearance model.Afterwards, the problem of predicting how the color signal arising in response to light reflected from the surface of an object is likely to change when the lighting alters is investigated. A new method, called the Gaussian Metamer (GM) method is proposed for predicting what a color signal observed from a surface under a first light is likely to be when the same surface is lit instead by a second light. Due to metamer mismatching, there is not a unique answer for this problem. Our approach is to use one of the possible metamers that is likely to do well on average. The results outperform other state-of-the-art prediction methods.