Compare Gray World and White Patch white balance.

What's really being asked

The interviewer wants to see that you know the foundational color-constancy heuristics, exactly what each assumes, when each breaks, and why data-driven methods help.

The full answer

Gray World assumes that, averaged over the whole image, the scene is achromatic; you compute the mean of each color channel and apply gains so the means become equal, neutralizing the illuminant's color cast. Its failure mode is a scene dominated by one hue, like a close-up of grass or a red wall, where the average is genuinely not gray, so it over-corrects. White Patch, or max-RGB, assumes the brightest patch corresponds to a white reflector under the illuminant; you take the per-channel maximum and scale so those maxima align. It fails when highlights are clipped, when the brightest object is colored, or with sensor saturation. A learning-based method instead trains on images with known ground-truth illuminants and predicts the illuminant color from richer features, capturing semantic and statistical priors that fixed heuristics cannot, and it degrades more gracefully on hard scenes.

The mistakes people make

Stating the algorithms without their assumptions, or claiming they always work. Confusing white balance with overall exposure correction. Forgetting that both classic methods are single global estimates.

What usually comes next

Shades of Gray and Gray Edge generalizations, how illuminant ground truth is captured with a calibration target, multi-illuminant scenes, and metrics like angular error.

A concrete example

Under warm tungsten light, Gray World raises the blue gain because the average is too red, while White Patch picks the brightest pixel as the reference white; on a photo that is mostly an orange sunset, Gray World wrongly drains the warmth, whereas a learned model recognizes the sunset and keeps it.