Machine colour constancy: a work in progress

The term “colour constancy” denotes the human capability to recognise an object as the same entity even if its colours change due to illumination. Implementing such a human capability is highly desirable in computer vision to retrieve, detect, recognise and track objects regardless of the illumination. Many efforts have been made in this direction, leading to several “machine colour constancy” (MCC) algorithms, that is, routines inspired by human colour constancy (HCC) that are aiming to achieve an image representation independent of the light and thus invariant against light changes. Different to HCC, which is subject to illusions and only partially removes illuminant effects, MCC pursues the implementation of a perfect, total removal of light effects from images. Such an implementation represents a major challenge and is therefore a focus of ongoing research. In fact, the currently available MCC algorithms only work under constrained domains. Therefore, a priori knowledge about the image content is needed to choose the most appropriate MCC procedure and to properly set its parameters. Here, we present the main common assumptions underlying most of MCC algorithms. Our work shows that two main issues should be addressed in the future to ensure MCC is more efficacious and useable: first, the lack of a reliable MCC algorithm operating without supervision in multiple scenarios; and second, the difficulty which any user encounters when choosing the MCC procedure most appropriate for a particular application.