We start this chapter by recalling some basic notions of white balance that were presented in Chapter 1. The white balance process aims at emulating the color constancy ability of the human visual system and consists of two steps, illumination estimation and color correction. Ideally it should be performed not on-camera but afterwards, as oﬄine post-processing: the rationale, as with demosaicking and denoising algorithms, is that with oﬄine postprocessing we have much more freedom in terms of what we can do, not being limited by the constraints of on-camera signal processing (in terms of speed, algorithm complexity and so on). Oﬄine we may use more sophisticated color constancy methods, which do not require all the simplifying assumptions used in practice for real-time processing. But oﬄine white balance requires that the image data is stored as raw, i.e. as it comes from the sensor, and many cameras do not have this option, storing the image frames already in color corrected (and demosaicked, and compressed) form. If oﬄine white balance is not an option, then the next best possibility is that of manual illumination estimation: the operator points the camera towards a reference white, such as a simple sheet of paper, and the triplet of values recorded for this object (at, say, the center of the image) is used by the camera as an estimate of the color of the illuminant. This method ensures that objects perceived as white at the scene will also appear white in the recorded images, and in general that all achromatic objects will appear gray. But cinematographers often find this effect too realistic, and use instead the manual white balance for artistic expression by fooling the camera, giving it as reference white an object with a certain color; in this way, by deliberately performing a wrong color correction, a certain artistic effect can be achieved . Finally there is the option of automatic white balance (AWB), where the illumination estimation is automatically performed on-camera and the color correction (an approximation of “discounting the illuminant”) is carried out on the raw data domain or just after color interpolation .