Browsing by Author "Santos, Fernando Lamar Corrêa dos"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Predicting video memorability using traditional and incremental approachesPublication . Santos, Fernando Lamar Corrêa dos; Sabino, André Miguel Guedelha; Estima, Jacinto Paulo SimõesThis study investigates the viability of incremental training as an alternative to traditional training methods for video memorability prediction, particularly in hardware-constrained environments. Using the ViViT model, a transformer-based architecture, the research seeks to address the primary question of whether incremental training can provide stable and consistent performance with reduced computational demands (RQ1). Two experiments were conducted: one comparing incremental and traditional training methods and another applying incremental training to the full dataset. The results indicate that incremental training is a feasible alternative, offering comparable performance in error metrics such as Mean Squared Error (MSE) and Mean Absolute Error (MAE), while significantly reducing the computational load. However, incremental training exhibited limitations in ranking accuracy, as measured by Spearman’s Rank Correlation (SRC), when compared to traditional methods. The findings suggest that incremental training can provide a practical solution for video memorability prediction in resource-constrained scenarios, but further refinement is needed to improve its performance in rank-order tasks. Future work should explore architectural optimizations, optimizing input configurations, expanding datasets, incorporating multimodal data, and tuning the ViViT architecture for better long-range dependency handling.