MVR Journal Club

This Wednesday, Dr Nadja Schinkel-Bielefeld, from the Department of Brain and Cognitive Sciences, MIT will give a presentation titled: " Connections Underlying Human Contour Integration are Adapted to Natural Image Statistics" (abstract below). Dr Schinkel-Beilefeld is an invited speaker, who is being hosted by Dr Kathy Mullen

Where: Royal Victoria Hospital, Room H4-14

When: Wednesday, 18th of February 2009, 1PM

Abstract:Humans are very efficient in detecting even jittered or partly occluded contours. But which mechanisms are capable of explaining this ability? A large class of contour integration models employs a so-called association field (AF), which reflects the connection structure between neuronal units representing different edge elements, for linking localized edges to a global contour. Here we investigate systematically how such an AF should be constructed to explain psychophysical contour detection data, combining ideas from theory, experiment and natural image statistics. We tested AFs of different symmetry and geometry in a Bayesian contour integration model, comparing it to human contour detection. Our psychophysical experiments revealed that correlations between human decisions are much higher than expected from their detection performances alone. This suggests that different humans make identical errors, voting for the same illusory contours in the stimulus. Hence we require our contour integration model not only to reach human performance, but also to reproduce these correlations and their characteristic time course. From all the AF choices we explored, a connection structure extracted from natural image statistics explains human contour integration best. From this we conclude that the human visual system is optimized for recognizing contours in natural images.