Xiangteng (Kai) Zhao¹ (), Howard Jia He Tan¹, Alejandro Lleras¹, Simona Buetti¹; ¹The University of Illinois Urbana-Champaign

Visual salience is thought to be the key determinant of eye movements under free-viewing conditions. The MIT Salience Benchmark is a resource that compares the success of computational toolboxes at predicting salience (and thus, eye movements) in digital images. However, a close look at the MIT Benchmark paper reveals that investigators used “memorization” instructions rather than a true “free viewing” instruction, when collecting the ground truth eye movement data. It is less clear how salience ought to impact eye movements under memorization conditions. The impact of the memorization instructions on eye movements could be problematic if subjects’ eye movements are different under the two instruction conditions. We hypothesized this would be particularly pronounced when indexing eye movements to low salient objects because whereas in free-viewing conditions, participants have little reason to visit these locations, under memorization conditions, participants might feel a need to visit those locations to better commit that information into working memory. The current study explored how eye movement behaviors towards low salience stimuli differ as a function of these two instruction conditions. Subjects were exposed to 96 varied scenes across two blocks of trials, engaging in either a memorization or a free-viewing task. Eye gaze was tracked using an Eyelink1000 Plus and fixations were categorized by the item they landed closest to. Results across both conditions revealed significant task-driven differences in fixation selectivity. A 13% increase in fixation count and a 42% increase in fixation duration on low salience objects were observed during memory tasks, indicating that memory instructions tend to produce fixation maps that overemphasize low salience regions of the scenes. These results question the validity of the MIT Salience Benchmark rankings. Further examinations and research are needed to validate these findings.

Acknowledgements: This material is based upon work supported by the National Science Foundation under Grant No BCS1921735 to SB.