Abstract
Purpose: Extensive research into the architecture of human scene perception and human figure-ground segmentation show that both local and configural processes play a role. Local factors include bottom-up edge segmentation enabling small regions to be fused into figural regions. Configural factors include top-down processes such as grouping and meaningfulness. Barghout (2009, 2011) suggested a natural-scene-perception architecture comprised of nested hierarchies of "spatial taxons" with the rank-frequency distribution predicted by a law of least effort, where attentional resources were minimized and utility optimized. Because computer vision models often provide insight into human perception, we decided to build a computer vision segmentation model that used spatial-taxon designation as a "meaningfulness" configural cue. Methods: The computer model used fuzzy-logic inference to simulate low-level visual processes and few rules of figure-ground perceptual organization. The model was required to conform to a spatial-taxon’s "meaningfulness" cue. We collected 70 real images composed of three "generic scene types", each of which required a different combination of the perceptual organization rules built into our model. We then used our model to segment the generic scene types. Two human subjects rated image-segmentation quality on a scale from 1 to 5 (5 being the best). Results: The majority of generic-scene-type image segmentations received a score of 4 or 5 (very good, perfect). ROC plots show that this model performs better on generic-scene-type images than normalized-cut ((Martin, Fowlkes, Tal, and Malik (2001).
Meeting abstract presented at VSS 2013