Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This study evaluates the performance of leading computer vision saliency models against real-world gaze data from 3,023 adults, revealing that a simple untrained central marker outperforms all trained models. The findings indicate that the additional features incorporated by these models do not align with actual viewer attention, resulting in systematic biases favoring certain demographic groups. By proposing a new framework for assessing model performance based on gaze data, this research highlights the need for inclusive design in attention prediction systems.
An untrained central marker outperformed advanced saliency models, exposing significant biases in how these systems predict viewer attention across demographic groups.
Computer vision saliency models predict where people will look, one map per image, and a billion-dollar predicted-attention industry sells those maps in place of measuring real viewers. I test the leading models from the audience side, against 11.4 million webcam gaze points from 3,023 US adults recruited to national quotas, viewing circulating news photographs. I show that an untrained central marker outperforms every trained network, because the content the networks add on top of the center falls where these audiences never look. What accuracy remains is systematically biased, favoring younger, White, and moderate viewers over older, Black, and ideologically extreme ones. I propose a way forward and build on what a group's own gaze reveals about whether a model can learn that group at all, and I apply it across every demographic axis this sample supports. Ultimately, I show how systems that decide what people see can learn to see everyone, and this study supplies the standard by which such a claim should be judged.