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

Poster C97 in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Humans Remember People; Models Remember Pixels: Exposing Blind Spots in Visual Memorability Prediction

Benjamin TenWolde¹, Virginia R. de Sa¹; ¹University of California, San Diego

Presenter: Virginia R. de Sa

Memorability is highly consistent across viewers, yet the models that predict it remain poorly understood. We analyze two architectures (ResMem, ViTMem) using attribution analysis, synthetic probes, and counterfactual interventions on THINGS. Both show little sensitivity to person presence (r=0.145 for humans vs. r=-0.068 for ResMem) and instead respond to chromatic and frequency statistics not linked to behavioral evidence. Gradient-guided pixel perturbations invisible to humans (L_∞<2/255) shift predictions by ∼0.56, more than eight times the person-presence effect, indicating that these models rely on pixel-level statistics rather than the scene-level processing that underlies human memorability.

Topic Area: Computational Models of Vision & Visual Cortex

Extended Abstract: Full Text PDF