Abstract
Natural materials such as skin, soap, jade, and wax, look translucent because light scatters inside them. Though humans are skilled at discriminating subtle differences in translucent appearance, little is known about translucency perception. We study how translucency perception is related to scattering parameters and object geometry, by measuring perceptual similarities between images rendered with varying scattering parameters, and using these similarities to learn low dimensional embeddings of the images. We compared embeddings for color and gray-scale images in settings where objects were only partly visible. Images of a dragon model were rendered, in gray-scale and color, using 16 different parameters from the spherical polydispersion scattering model, under constant natural illumination. Each observer was shown 1680 ordered triplets of images, and asked to indicate whether the center image was more similar to the left or right image. These paired-comparisons were used with a non-metric multidimensional scaling method to learn 2D embeddings of the data. The 2D embeddings from four observers resembled a U-shaped curve, with images consistently ordered from opaque to translucent. To isolate the effect of different cues in the scene, we repeated our experiments using only images from the opaque side of the curve, masking out either the top- (specular highlights) or bottom-half (see-through effect) of the images with a semi-transparent layer. For color images, the embedding of bottom- and top-masked images had the same shape as under the full-view conditions, while the top-half masked ones had different image ordering. For gray-scale images, the embedding of bottom-half masked images did not have clear structure, while that of top-half masked images was U-shaped but with different orderings from the color case. This suggests that translucency perception depends on shape and there is an interaction between color and shape. The effect of shape on translucency is more significant for gray-scale objects.
Meeting abstract presented at VSS 2012