Metamerism in Spatial Vision

Lewis D. Griffin

Medical Imaging Science IRG,

King’s College, London

Metamerism effects spatial vision i.e. for any given pattern of responses of visual neurons, there is a large metamery class of images that could have caused it. The term ‘feature detection’ acknowledges the belief that a visual system should somehow proceed from initial quantitative measurements (e.g. linear neuron responses) to categorical qualitative descriptors (e.g. dark blob on the light side of a convex edge). This step is complex even when applied to single images, so the prospect of developing a theory that applies to vast metamery classes of possible images is dismaying. Koenderink [1] has pioneered a strategy that solves both problems: identifying a unique simple iconic image within each metamery class. This dispenses with the need to deal with the full metamery class, and since the iconic images are simple the extraction of qualitative structure will be facilitated.

The correct rule for selecting from a metamery class remains unknown, though candidates have been explored [2, 3]. We have examined six new candidate rules based upon minimizing norms of the luminance or the gradient magnitude. We have also explored a further rule based on choosing the most likely (ML) explanation within a metamery class. This we approach by calculating (on the basis of a natural image collection) the most likely explanation of 1^st order structure in 1D and 2D. Finally, we consider to what degree there is coincidence between the norm and ML approaches. We find such coincidence plausible though our best candidate (the total variation norm) has clear shortcomings.

[1] Koenderink JJ (1993) What is a feature?, J. Intell. Syst. 3(1):49-82.

[2] Tagliati E & Griffin LD (2001) Features in Scale Space: Progress on the 2D 2nd Order Jet, In: Scale-Space and Morphology in Computer Vision vol. 2106 (M Kerckhove, ed.), Springer, pp. 51-62.

[3] Koenderink JJ & van Doorn AJ (1996) Metamerism in complete sets of image operators, In: Advances in Image Understanding '96, pp. 113-129.