IO: Towards an Alternative Image of Human Vision

Abstract

IO is a system which interprets images of geometric objects with straight edges. It was prompted by examples where human vision behaves unlike the well known ‘blocks world’ programs of Huffman, Clowes, Waltz and others. Some of these examples are described. IO captures many of these distinctively human behaviours by reversing the strategy of the ‘blocks world’ programs: early processing is driven by the expectation that edges will form geometrically regular frameworks, and only at the end is it asked how surfaces are connected. IO begins by finding clusters of junctions which are close to each other and exhibit a regular geometric pattern in 2-d (image) space. Edges are then assigned 3-d slopes using a rectangularity assumption in each cluster. Finally IO considers the way surfaces meet at edges, using its knowledge about 3-d slopes to constrain its decisions. Restricted types of information pass between analyses concerned with different clusters. This allows IO to mimic a striking feature of human vision: a high level of consistency is generally maintained, but inconsistencies arise and go unnoticed in the pictures which people see as Impossible Objects. IO’s approach lends itself to parallel implementation in several respects, and the early parts of the system are now implemented in OCCAM.KeywordsLine DrawingHuman VisionOrientation ProfileEdge OrientationImpossible ObjectThese keywords were added by machine and not by the authors. This process is experimental and the keywords may be updated as the learning algorithm improves.