Local Surface Patch Classification Using Multilinear PCA+LDA on High-Order Image Structures Compared to Human Observers

摘要

What information do we use to determine the curvatures of local surface patches? In a 5-AFC decision task, observers judged the curvatures of local surface patches viewed through an aperture, including Bells, Dimples, Furrows, Humps, and Saddles that were cylindrically projected onto a sphere. Numerous high-order image structures were computed from stimulus luminance values. Classical and Multilinear PCA were performed on image structures, which were dimensionally degraded until mean model performances (i.e., proportions of correct discriminations) matched the mean performance of observers. The posterior probability distributions of the LDA classifiers were then correlated with human error confusions. Among the image structures that were examined, the strongest predictors of human performance involved 2nd-order derivatives of the luminance patterns. Using more than one image structure at a time did not reliably improve model prediction, leading us to choose Laplacian of Gaussian arrays and Multilinear PCA+LDA for further analysis. The model accounted for approximately 33% of the error confusions that were predicted by independent human observers. In other words, the model was about 1/3 as reliable as the test-retest reliability of independent human observations. It appears as though humans may use information analogous to high-order image structures to judge local surface contours, however the exact information guiding our perceptual judgments remains uncertain.