Image memorability using diverse visual features and soft attention

Abstract

In this paper we present a method for still image memorability estimation. The proposed solution exploits feature maps extracted from two Convolutional Neural Networks pre-trained for object recognition and memorability estimation respectively. The feature maps are then enhanced using a soft attention mechanism in order to let the model focus on highly informative image regions for memorability estimation. Results achieved on a benchmark dataset demonstrate the effectiveness of the proposed method.