Colour constancy in natural scenes: a computational look
Place: Large Lecture Room - CVC
Affiliation: Institute of Neuroscience, Newcastle University, UK
Colour plays an important role in human and computer vision, for image segmentation, material identification and object recognition.
Colour constancy , the phenomenon by which object colours remain stable under changing illumination, is vital for colour to be a reliable indicator of material or object identity.
Computationally, colour constancy is the formidable problem of extracting the illumination spectrum and recovering invariant surface colour from the varying image data. Accurate recovery is not possible without imposing constraints on scene formation, such as the existence of a single, spatially uniform light source, or identification of a white reference surface in the scene.
In this talk, I will review features of natural scenes which present both obstacles and potential solutions to the computational problem of colour constancy. In particular, the “polychromaticity” of natural surfaces , such as mottled bananas, may provide additional information about the light source, allowing familiar objects to act as reference surfaces to calibrate colour constancy. This use of familiar objects provides an explanation for the possible role of “memory colour” in colour constancy, first mooted by Hering in the 19th century.