Catalogue of Artificial Intelligence Techniques
Author(s): Jon Mayhew
The term was first used by Barrow and Tenenbaum (l983) to refer to a registered `stack' of retinotopic maps each of which makes explicit the value of a certain property `intrinsic' to the surfaces in the scene. The intrinsic images all have the same viewer centred coordinate system but carry information about different surface properties, such as: surface orientation, depth, reflectance, colour, texture and Optical Flow (this last intrinsic image describes the instantaneous velocity flow field in the scene). The computation of the intrinsic images is non-trivial, many of them are underdetermined when considered independently but global consistency constraints can be cooperatively exploited, e.g., surface boundary information carried by the reflectance image can be used to constrain the computation of the surface orientation from shading information. Over the last five years there have been considerable advances in understanding of the problems of computing the different intrinsic images. Their particular importance is that they are a vital stage in the computation of a representation intermediate between the lowest levels of image processing, whose descriptions are essentially 2D pictorial or iconic descriptions of the scene, and the higher levels of processing which describe the shapes of objects in terms of an viewer independent Object-centred Coordinate system. The intrinsic images are the first representation at which information concerning the 3D structure of the scene is made explicit. Object recognition schemes that have attempted to recover 3D shape descriptions directly from 2D shape descriptions have, to put it mildly, struggled. See also 2½-D Sketch.
- Brady, M., Computational approaches to image understanding, Machine Vision: The Advent of Intelligent Robotics (Brady, M., ed.), Addison-Wesley, Wokingham, 1986, pp.7--66.
- Barrow, H.G. and Tenenbaum, J.M., Computational vision, Proceedings of IJCAI-83, 1983, pp.39--74.