Abstract
Purpose: Prior theories of the perception of partially occluded figures have stressed bottom-up, local processing, in which depth cues are necessary for discriminating contours “extrinsic” to an occluded figure from those “intrinsic” to it. Results of our experiments reported previously indicate that any contour feature can be used in such discrimination (e.g., orientation, curvature, spatial scale). Further, the percept is not based on mere integration of visible contours of the figure, suggesting the operation of top-down, global processing. To verify these claims, we formulated and tested a new computational model.
Methods: The model is based on an exponential-pyramid architecture and processing involves two stages. The first stage (bottom-up) computes local variance of each contour feature (orientation, length) and verifies whether this variance is approximately constant across the image and spatial scale (receptive field size). Non-constancy of the variance indicates the presence (and position) of a figure in the image. The second stage (top-down) uses this statistical information to discriminate between intrinsic and extrinsic contours. The model has only one free parameter: the standard deviation of decisional noise. Model and human performance were compared across 11 experimental conditions.
Results: Subjects' performance was close to perfect when there was no overlap between the histograms of a givenfeature for the figure and for the occluder. Performance systematically deteriorated when overlap between the two histograms increased. The new computational model accounts well for all 11 experimental conditions.
Conclusions: Perception of occluded figures is not a purely bottom-up process integrating spatially local pieces of information. Instead, it begins with determining spatially global properties of the image, which are then used to make local perceptual decisions. Such a two-stage mechanism is computationally more robust.