VSS 2015 Winners > Hideki Tamura
Can the classifier trained to separate surface texture from specular shading infer geometric consistency of specular highlight?
Hideki Tamura1, Shigeki Nakauchi1; 1Department of Computer Science and Engineering, Toyohashi University of Technology
Separating the surface texture (changes in albedo) from surface shading is a fundamental task for the visual system (Kim et al., 2014). We investigated whether this task is possible only from the image statistics. Images of randomly generated “bumpy” surfaces with/without specular reflections were rendered as gloss/matte surfaces. Half of the images in each category were modified so as to have surface texture (pigmentation) by multiplicatively changing the intensities with binary noise patterns. These images of four categories (matte, gloss, textured-matte and textured-gloss; 1,800 images for each category) were normalized to have the same mean and variance of pixel intensity, and then 744 of PS statistics (Portilla & Simoncelli, 2000) were calculated for them. Next, we developed a statistical model to classify the images into four categories using their PS statistics by canonical discriminant analysis (CDA). It turned out that the trained classifier is able to clearly categorize the generated images, and the first two canonical coordinates correspond to gloss/matte and uniform/textured axes, respectively. With the trained classifier, we examined the hypothesis: computations for evaluating the geometric consistency of highlights on glossy surfaces are common with those for separation of changes in albedo from surface shading. Diffused matte components and specular highlight components were separated from original glossy surface images, and as the test images the matte surfaces were textured with the separated highlight components in the same manner as in the training “textured” image generation with various multiplicative factors (1.2 to 5.0). This pigmentation was applied to consistent and inconsistent locations (original highlight components were horizontally flipped) of surfaces (3,600 in total). The trained classifier was able to differentiate consistent/inconsistent categories, except when the highlight-shaped texture regions have very high or very low local contrasts, suggesting that inference of geometric consistency of specular highlights is partially done by albedo-shading computations.