Minimizing Worst-case Errors for Illumination Estimation

An image of a scene depends on the reflectance properties of the objects in the scene and on the illuminant incident upon those objects. Colour constancy is an ability to determine some properties related to or implied by an object’s reflectance regardless of the illuminant illuminating the scene. One of the main approaches to colour constancy relies on illumination estimation. While many illumination estimation algorithms exhibit excellent performance in terms of median error, the performance in the worst-case scenario remains quite poor. This thesis focuses on reducing the maximum errors. Several options were investigated and show promising results. These include: combining a variety of clues, multispectral imaging and “outlier” reduction. Combining a variety of clues has been explored in the past with a focus of minimization of median error. This thesis explores avenues that lead to lowering of worst case or high percentile errors. Additional information present in multispectral images may influence the performance of illumination estimation algorithms. The performance of illumination estimation algorithms for multispectral images was investigated and a new colour constancy algorithm that uses multispectral images was developed. It is possible that a single strong clue is wrongly influenced by a small part of an image. An algorithm is presented that splits the image into many parts, runs an illumination estimation algorithm on each of the smaller parts and combines the results in a manner that is less sensitive to outliers.