Contributed talk at ECVP !
in Conferences
An ideal observer model for grouping and contour integration in natural images. Check the slides of the talk here. Joint work with Pascal Mamassian and Ruben Coen-Cagli
Spatial context in images modulates visual perception reflecting optimization to the statistics of natural images. Popular models based on divisive normalization (ratio between target and context features) offer a link between optimal coding principles and contextual modulation in cortex. Here, we apply this framework to perceptual grouping in natural images, and more specifically to contour integration. First, we show that a successful model of image statistics based on normalization (termed Gaussian Scale Mixture, GSM) learns dependencies between colinear features in natural images. We then show analytically that image regions with strong normalization are statistical outliers of the model, and use this insight to distinguish high-salience regions due to high contrast from those corresponding to contours. To this aim, we extend the model to a mixture of GSMs, with each component encoding a specific orientation and curvature. Our model performs competitively on contour detection in natural images from the BSD500 database. We further evaluate our model by comparing its performance on iso-oriented Gabor elements embedded in Gabor noise (‘snakes’) to the psychophysics literature. Our model thus serves as an ideal observer for contour integration and a basis for future experiments on natural image segmentation.
