Towards a New Approach to Vision: Applications to Robots and Humans

Abstract

A description of a novel and very speculative approach to new research directions for human vision with application to robotic vision is described. The goal of the approach is to propose a plausible, implementable, spatial perception model for human vision and apply this model to a stereo robot vision system. The model is based on computer algorithms variously called “Multidimensional Scaling”, well known in psychology and sociology but relatively unknown in engineering. These algorithms can reconstruct a spatially accurate model to a high level of metric precision of a “configuration of points” from low quality, error prone non-metric data about the configuration. ALSCAL – a general purpose computer package adaptable for this purpose is being presently evaluated. This is a departure from typical engineering approaches which are directed toward gathering a low volume of highly precise referenced data about the positions of selected points in the visual scene and substitutes instead an approach of gathering a high volume of very low precision relative data about the interpoint spacings. It would seem that the latter approach is the one actually used by the human vision system.

The results are highly encouraging in that the agreement between test configurations of two and three dimensional configurations of points are very faithfully reconstructed from as low as 10 points in a configuration using only rank ordered (i.e. nonmetric) information about interpoint spacings. The reconstructions are remarkably robust even under human-like “fuzzy” imprecision in visual measurements.