Kobus Barnard, Brian Funt, and Lindsay Martin, "Colour Constancy Meets Colour Indexing," Submitted for publication, October 9, 2000. [ Full text (gzipped postscript) ]
In this paper we investigate the ability of colour constancy processing to improve illumination sensitive machine recognition tasks. So far, colour constancy results have been limited to simple error metrics which do not indicate how well current methods are suited to the often quoted application of illumination invariant object recognition. Thus in this work we set up a colour sensitive recognition task to be undertaken under the range of commonly encountered illumination. We present results relating recognition performance to computational colour constancy error. A comprehensive sampling of algorithms is used. The results are based on 220 images which have been made available on the on the internet (http://www.cs.sfu.ca/~colour/data).
Keywords: colour, colour constancy, colour constancy algorithms, indexing, object recognition, image data
Full text (gzipped postscript)