Year: 2026
Type: preprint
Abstract: Introduction: This study investigates the impact of modular components influencing image memorability, defined as a person's ability to recall previously seen images or objects. Although memorability is influenced by various elements, the extent to which these factors contribute to overall memorability remains underexplored. Methods: Existing models primarily focus on low-frequency information in visual scenes, often overlooking high-frequency data that may also capture traits contributing to image memorability. To address this, we introduce a novel Multi-Frequency Filter Bank Module (MFFBM) tailored for Graph Convolutional Networks (GCNs), enabling the extraction and utilization of low- and high-frequency information from graph representations of images and objects. Additionally, we explore multi-modal data sources such as visual saliency and image captions. Building on these components, we propose the Graph Frequency Filtering Multi-Modal Memorability (GRAFF-MM) model, which integrates MFFBM with multi-modal data for enhanced memorability prediction. Results: We evaluate our model on four benchmark datasets: LaMeM, SUN Memorability, OBJ Mem, and FIGRIM. The proposed model achieves competitive Spearman's rank correlations of 0.692 on LaMeM, 0.711 on SUN Memorability, 0.653 on OBJ Mem, and 0.676 on FIGRIM, demonstrating its effectiveness in predicting image memorability. Conclusion: We conduct an extensive ablation study to validate the effectiveness of the modules. We also visualize the learned features and edge cases to explain how our modules work in tandem to predict memorability.
Source: Research Square
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7222113386