Theory 08/Vision & Ocular Physiology /Visual perception/3 min
Colour Constancy Is a Prior, Not a Computation
The brain does not solve for the light source. It bets on what light sources are usually like.
Colour constancy works by applying a statistical prior over natural illuminants rather than by estimating the illuminant from the image, so it fails predictably in lighting that violates the prior rather than in lighting that is merely unusual.
01The evidence
- Colour constancy has been modelled as Bayesian inference over scene statistics, with the prior over illuminants doing the work that an explicit estimate would do.Bayesian colour constancy
- Bayesian models of human colour constancy have been tested against observer data, which makes the prior a measurable object rather than a metaphor.Bayesian model of colour constancy
02The tension
The dominant account treats constancy as an inverse problem: the visual system estimates the illuminant and discounts it. That predicts failures when the estimate is hard. In practice, failures are not predicted by difficulty but by whether the scene resembles natural conditions.
03The theory
Natural illuminants have a characteristic distribution: daylight follows a locus, incandescent light clusters elsewhere, and mixtures are constrained. The visual system has learned that distribution and uses it as a prior.
Constancy then becomes inference, not estimation. The percept is the posterior, which is dominated by the prior when the image is ambiguous. That predicts failure when the actual illuminant lies outside the learned distribution, regardless of how much information the image contains.
The prediction is a dissociation: scenes with plenty of cues but unnatural illumination should fail, while scenes with few cues but natural illumination should succeed. That is the opposite of what an estimation model predicts, and it is directly testable.
The applied consequence is that display calibration and colour grading should consider the prior, not just the physical light, because an accurate simulation of an unnatural source will be misperceived in a predictable way.
04What this is not
Those models estimate the illuminant and discount it, predicting failure when estimation is hard. This predicts failure when the illumination is improbable, which is a different and separable variable.
05Predictions
- P1Constancy should fail for unnatural illuminants even when many cues are available.
- P2Constancy should succeed for natural illuminants with few cues.
- P3Observers' priors should be measurable and should match the statistics of their visual environment.
06What would refute it
Constancy tracking cue availability rather than illuminant probability would restore the estimation account.
07What it would change
It would change how colour is managed in displays and in imaging, and give a measurable prior for observers.
08Evidence base
- Colour constancy as Bayesian inference on scene statisticshttps://researchmap.jp/toyota-toshihiro/presentations/15623398?lang=en
- Bayesian model of human colour constancyhttps://scholar.google.com.eg/scholar?btnI=1&cluster=3900168027893007327