The Chromagenic Colour Camera and Illuminant Estimation

A chromagenic camera takes two pictures of each scene. The first is taken as normal but a specially chosen coloured filter is placed in front of the camera when capturing the second image. The chromagenic filter is chosen so that the combined image makes colour constancy, or white point estimation easier to solve. The chromagenic illuminant estimation algorithm is very simple. We compute the expect relations, currently implemented as 3x3 matrix transforms, between unfiltered and filtered RGBs for a range of typical lights. These relations are tested in situ for a given chromagenic image and the one that best predicts the image data is used to designate the illuminant colour. However, in experiments we found that a 3x3 matrix transform, while generally quite accurate, can fail to model the relationship between filtered and unfiltered RGBs for some colours (e.g. saturated colours) and so, the chromagenic algorithm which works very well on average can nevertheless, on occasion, work poorly. In this paper we assume that convex combinations in local areas of RGB space are translated to the same convex combinations for corresponding filtered RGBs and use this insight to relate filtered and unfiltered RGBs. These locally convex relations model the image data more accurately. Testing these relations in situ in images and choosing the one which best models the data provides surprisingly effective illuminant estimation algorithm. Experiments demonstrate that the chromagenic colour constancy algorithm provides superior illuminant estimation compared with conventional approaches (Gamut mapping, color by correlation, max RGB etc). This result holds across many different data sets. The method is also demonstrated to work on real images. The plausibility of the chromagenic approach for human vision is also discussed.