Early Scene Analysis: Rapid Processing of Contours Surfaces, and Objects in Human Vision (AASERT)

摘要

This report covers the AASERT grant which is a companion grant to AFOSR 94-0189 of the same title. Over the duration of the grant, support and training were provided for 5 different graduate students and 7 undergraduates. The work concentrated on how 2-D information is built up from the parallel analysis of a set of visual attributes and how this information contacts memory in order to construct 3-D representations of the visual scene. We were specifically interested in the early stages of these processes which may operate under simplified assumptions in order to gain speed. We examined the rapid decomposition of image values into object features (reflectance, transparency, orientation, 3D position) and illumination features (shadows, shading, highlights). We also evaluated the nature of the representation achieved at these early levels and whether the output of each of the many early stages is independently available to higher level processes or only one final 'result' gets passed along. Finally, we studied the initial contact between the image contours and memory in recognition.