Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Memorability Varies with How Visual Information Is Sampled by Humans and Models

Sule Tasliyurt Celebi¹, Katharina Dobs¹; ¹Justus Liebig Universität Gießen

Presenter: Sule Tasliyurt Celebi

Humans consistently remember some images better than others—a phenomenon known as memorability. Computational models such as ResMem can predict image memorability from visual features, yet it remains unclear whether these predictions rely on the same spatial prioritization as humans. Here, we tested whether model–behavior agreement depends on shared attentional selection. Participants performed an old–new memory task with eye tracking on natural scene images containing faces, text, both, or neither, selected based on high or low ResMem scores. While ResMem distinguished high‑ from low‑memorability images overall, its performance varied across conditions and failed for text‑only images. Critically, model–behavior discrepancies showed a non-monotonic relationship with gaze-saliency alignment, with best prediction at intermediate levels of alignment. However, this relationship disappeared for images containing both faces and text, where predictions remained accurate irrespective of gaze–saliency alignment. These findings suggest that memorability depends not only on image features but also on how visual information is selectively sampled, highlighting the role of shared spatial prioritization in model–behavior alignment.

Topic Area: Computational Models of Vision & Visual Cortex