Illumination-invariant representation for natural colour images through SIFT matching

Illumination conditions may severely impair many computer vision applications. Defining an illumination-invariant representation from a colour image is therefore of great interest. Assuming that two views of a scene are available, we show that it is easy to derive such a representation from correspondences between interest points, within a Lambertian reflectance / Planckian lighting model. An application to shadow removal or lightening is also discussed.