McDermott, Joshua Hartmann; (2002) Junction detection: Computation and psychophysics. Masters thesis (M.Phil), UCL (University College London).

Text Junction_detection_Computatio.pdf Download (10MB)

Abstract

Junctions, formed at the intersection of image contours, have been thought to play an important and early role in vision since the time of Helmholtz, and have recently been the focus of much research in vision science. The interest injunctions can be attributed in part to the notion that they are local image features that are easy to detect which nonetheless provide valuable information about important scene events like occlusion and transparency. Despite the prevalence of this view, the extent to which it holds for junctions in real images remains unclear. In this thesis we put this notion to the test. Our first approach involves psychophysical experiments with real images as stimuli in which we use the human visual system as a tool to measure the visual information available in local image regions. We test the ability of observers to detect occlusion points, which by hypothesis are signified by local junctions, given only a small image region surrounding the occlusion point. Human performance in our experiments places constraints on how junctions might be detected and what their role in vision might be. We also attempted to train neural networks to detect junctions given only a small image patch centered on the junction. Together, our experiments suggest that although some junctions are locally defined and can be detected with simple and local mechanisms, a substantial fraction necessitate the use of more complex and global processes, and may not play the bottom-up role in vision which they have often been ascribed.

Download activity - last month

Export as JSONCSVXML

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item