Centre de Visió per Computador - Universitat Autònoma de Barcelona

Why a device-independent dataset?

Almost all current image databases in Computer Vision are determined by the imaging device that gathered its component pictures. There are off course very good calibrated databases of hyperspectral images, but these tend to be very difficult and time consuming to produce and in consequence they usually consist of just a few examples. Other databases are more specialised (they are defined in some specific colour space) or they consist of some special kind of imagery (such as man-made objects under specific illuminants, etc). This lack of a device-independent, all-purpose and open database becomes an important drawback when we need very accurate, statistically representative depictions of the colours present in ordinary images. Our dataset of calibrated images seeks to close this gap. It contains two main improvements over previous databases: first, it has been gathered using a calibrated camera (whose calibration process is explained in this website) and second, it accommodates a grey ball of known reflectance on the lower-left side of the image. By calibrating the camera we aim at defining the chromatic properties of the light hitting each sensor unit (pixel) in a device-independent colour space such as the CIE1931 XYZ and by adding this ball, we expect the user to be able to infer the chromatic characteristics of the illuminant at the moment each picture was taken (so the dataset could be used in colour constancy studies or surface reflectance studies). The grey ball was attached to the camera by a metallic pole which allows us to vary the position and the size of the ball in the picture. We experimented with these parameters to try to position the ball so that it samples the light falling over the whole of the scene.

Why a ball and not a card? Most applications which require a sample of the illuminant use a flat grey card (such as the standard Kodak grey card or similar). There are two main problems with this: first the card has to face the illuminant in every circumstance, which is cumbersome and time consuming and second, it reinforces the illusion of a perfectly uniform illuminant. We know that in most circumstances, and certainly in natural conditions, part of the illumination comes from the sun and part comes from the sky or through inter-reflections (mostly in shaded areas). Having a ball does not solve this issue but at least it tells us when differences in the illuminant’s chromaticity become significant.