Our primary focus in the Vision Lab at SFU is in understanding colour: How are colours perceived? How can colours be reproduced accurately on different media? In what ways does colour help in understanding images? Understanding colour is a much more difficult problem than most people suspect. Often poor colour rendition results more from our limited understanding of colour perception than it does from limitations of our colour producing devices.

We subscribe to a computational view of colour; namely, that human perception and use of colour can be explored and explained as computations. The fundamental problem of colour is to explain how we see colours as relatively stable despite the fact that the light reflected into our eyes from an object varies dramatically with the light illuminating the object. Colour and computers have become much more intertwined in recent years as colour displays and colour printers have become more affordable. Since colour is a perceptual, not a physical quality, it is crucial to have a good model of how we perceive colour in complex environments if we are to get predictable results from these devices.

Colours are difficult to reproduce correctly, but why? While we've all experienced untrue colour while using home video cameras or viewing prints from our local photofinisher, now we have colour printers frustrating us with colours that look very little like the nice colours we previewed on our LCD display. When the colour doesn't look right, it's natural to feel that the printer and display are not calibrated properly--- and of course perhaps they're not--- but that's not the fundamental problem. The fundamental problem stems from the fact that colour reproduction, simply is not a matter of reproducing identical physical phenomena, as it is in the case of sound reproduction in which a similar pattern of sound waves is recreated, but rather a matter of creating perceptual equivalences.

For us to build machines that reproduce colours accurately or to make effective use of colour in robotics requires that we understand human colour perception; and the last decade has produced many interesting new computational theories of colour coming from both computer science and psychology. A central concern of these theories is to describe how colour depends or does not depend on the incident illumination. A coloured surface cannot be seen unless we shine some light on it, but the spectrum of the reflected light depends on the product of the spectrum of the incident light's spectrum and the surface's reflectance. It's natural to think of a surface's colour as a feature of the surface itself, but the spectrum of the light energy reaching the eye has the two factors of illumination and reflectance confounded into one. In order to determine the true surface properties, the effect of the illumination must be taken into account.

In the upper left is the image of a scene taken under an orangish, tungsten light which has the effect of turning the background overly green and the whites a bit yellow. In the bottom right is the target image of the same scene but under the standard illumination for which the camera is calibrated. Since we cannot always control the illumination, our goal is to correct automatically the colours in the input image so that they will look like those in the target image. (The fuzziness of the images is due to a high JPEG compression factor--- concentrate on the colours.)

The top right is a 'corrected' image produced by the standard grey-world colour balancing algorithm, which assumes that the average of all the colours in the scene is grey. In the bottom left is the much better result produced by our new, more sophisticated algorithm developed in the Vision Lab. As you can see, the result of our new algorithm is much closer to the target image than that produced by the grey-world algorithm.