Brahan Wassan-Aklilu¹, Ohad Ben-Shahar²; ¹Ben-Gurion University of the Negev

Observers often fixate near the center of images ("center bias"), but it is unclear which geometric "center" attracts early gaze. We address this with two eye-tracking experiments (EyeLink 1000, fixed head) using masked scenes from CAT2000. In the pilot Experiment 1 (N=20, 540 trials), each image was masked with a randomly generated star-shaped polygon having several spatially distinct centers. Trials began with a peripheral cue followed by 4's of free-viewing. Measuring mean distance to each candidate center and their average, we confirmed that the second-fixation (first post-cue) showed the strongest centrality bias that followed the polygon geometry rather than the monitor center or image content. Out of the different possible centers tested, four exerted the strongest and nearly identical mean distances: center of mass (CoM), bounding-box center (BBC), convex-hull center (CHC), and inscribed-circle center (ICC), which Experiment 2 examined more closely. Experiment 2 (N=25, 429 trials) was designed to resolve this ambiguity by jointly controlling image content and polygon geometry within a decision tree architecture. We selected 390 images from six CAT2000 categories spanning strong versus weak center bias (Object, Sketch, Fractal vs. Jumbled, Satellite, Indoor). Each image was masked with one of 13 evolutionarily optimized polygons, carefully structured to isolate and evaluate the predictive power of each of the four candidate geometric centers. Stimuli polygon set comprised a baseline of coincident centers, maximally separated-centers concave-shape, three pairwise-cluster configurations, and eight isolated-center configurations. The same polygons were also shown apictorially to measure purely geometric bias. Pooled results indicate that post-cue-fixations are closest to CoM, suggesting that center bias is dictated by the center of mass of the visible stimulus region. While being modulated but factors such as pictorial structure and semantics, these results provide a precise target for center priors in computational salience models.