Item

Abstract

Upon glimpsing at an image, we effortlessly perceive structures from within. What charac- terizes this process? Historically, gestalt psychologists have suggested that people tend to group nearby similar image parts together as a whole. Can an algorithm that partitions images into sub-parts based on similitude characterize visual per- ception behavior? We look at the normalized min-cut algorithm and its correlation to the recognition difficulty of image parts. The algorithm transfers an image seg- mentation problem into a graph-cutting problem, approxi- mating an energy optimization problem that preserves within-graph similarities. We study whether the number of computational steps needed for the algorithm to isolate an image part correlates with participants’ difficulty in recognizing that part, and whether higher exposure time correlates with further computational steps. We propose a psy- chophysics paradigm to study subjects’ recognition behavior upon seeing images tiled by amorphous sub- parts. We found that an increasing cut no. of image subpart (more computation steps) is harder for subjects to recognize after a brief exposure time, and longer exposure time increases the recognition ease, consistent with the model’s prediction that higher cut no. demands more computation steps to isolate. Our study relates the recognition difficulty of image parts with the computational resources needed to solve an optimization problem of grouping by similarity.