Abstract
Within the real world, surface reflectance varies over a very limited range. Illumination, however, can vary over many orders of magnitude. The luminance signal we receive from the environment is the product of these two variables, prompting the question; how does vision go about extracting the behaviourally important surface reflectance signal from the larger but less relevant illuminant? A possible solution could be to use a method of contrast adaptation whereby the incoming luminance signal is scaled based on the temporally averaged (log) input. Here we investigated; 1) how well this simple lightness constancy mechanism could work in the real world, 2) whether the optimal (log-linear) system for extracting reflectances has temporal parameters that match those observed behaviourally, 3) whether an exponential temporal weighting function (a Kalman filter) worked better than a power law impulse response, as has been suggested in the literature, and 4) whether the nature of a full Bayesian solution to this problem could shed light on how previous contrast exposure can change the temporal properties of contrast adaptation. To this end we measured, over 6 days, the illuminant incident on a person walking in a number of environments. Assuming three eye movements per second, and exposure to a distribution of reflectances measured within a forest environment, we were able to create luminance time series for which the actual series of reflectances that generated them were known. Using this data set we found that; 1) contrast adaptation greatly improves signal to noise ratio, 2) the best exponential time constant was 8 seconds and this matches the behavioural data, 3) an exponential weighting function is better than a power law; and 4) a full Bayesian solution functions better still, adaptively changing its time constant over a time scale of about 20 seconds.