Item

Abstract

It has been twenty years since David Marr produced his ground-breaking framework of vision as a hierarchical combination of distinct modules, each performing its own computation on retinal input. This modular theory is a computational simplification that treats the goal of vision as the extraction of visual cues. Researchers have been addressing how each of the modules could possibly operate in isolation. To this end we have had many ingenious inventions such as the random-dot stereogram, intricate plaid patterns ,and colourful Mondrians. However, the simplifications afforded by such thinking are often offset by the difficulties they introduce. First, the world does not consist of plaid patterns--it's more complex than that. Second, isolation of visual information almost inevitably leads to ambiguity in the reconstruction of the real world. The ill-posedness of vision with isolated cues can be resolved by the combination of cues: disparity, shading, texture, motion, etc. Using statistical methods such the Bayesian framework allows for the maximisation of the information derived from various sources. But, it seems still not to be enough. Perhaps a better way of thinking about seeing can be reformulated; vision does not start at the retina. Vision starts when a particular task has to be performed. The role of vision is not one of reconstruction of the real world in the brain but one of serving the needs of a mobile active being that functions in the real world. The talks presented in this session perhaps give a flavour of how it has been in Vision and also perhaps a flavour of how it will be in the future.