Color in Computer Vision: Fundamentals and Applications

Chapter 12: Evaluation of Color Constancy Methods

Evaluation of illuminant estimation algorithms requires images with a scene illuminant that is known (ground truth). ¹The general experimental setup is as follows. First, part of the data is used for training, if the algorithm requires this. Then, the color of the light source is estimated for every remaining image of the database and compared to the ground truth. Various publicly available data sets are discussed in Section 12.1. The comparison requires some similarity or distance measure, discussed in Section 12.2. Alternative setups exist, depending on the application. For instance, Funt et al. [219] describe an experiment to evaluate the usefulness of color constancy algorithms as a preprocessing step in object recognition. However, in this chapter, the intended application is correction of an input image for the color of the light source, that is, white balancing.

12.1 Data Sets

Two types of data that are used to evaluated color constancy methods can be distinguished: hyperspectral data and RGB images. ²Databases containing hyperspectral data sets are often smaller (less images) and contain less variation than data sets with RGB images. The main advantage of hyperspectral data is that many different illuminants can be used to realistically render the same scene under various light sources, and consequently a systematic evaluation of the methods is possible. However, the simulation of illuminants generally does not include real-world ...