Colour for Computer Vision: A Primer?

Still, much of computer vision is predicated on greyscale imagery. There are good reasons for this. For much of the development of computer vision greyscale images were all that was available and so techniques were developed for that medium. Equally, if a problem can be solved in greyscale - and many can be - then the added complexity of starting with 3 image planes as oppose to 1 is not needed. But, truthfully, colour is not used ubiquitously as there are some important concepts that need to be understood if colour is to be used correctly. In this chapter I summarise the basic model of colour image formation which teaches that the colours recorded by a camera depend equally on the colour of the prevailing light and the colour of objects in the scene. Building on this, some of the fundamental ideas of colorimetry are discussed in the context of colour correction: the process whereby acquired camera RGBs are mapped to the actual RGBs used to drive a display. Then, we discuss how we can remove colour bias due to illumination. Two methods are presented: we can solve for the colour of the light (colour constancy) or remove it through algebraic manipulation (illuminant invariance). Either approach is necessary if colour is to be used as a descriptor for problems such as recognition and tracking. The chapter also touches on aspects of human perception.