Bottom up Saliency Evaluation via Deep Features of State-of-the-Art Convolutional Neural Networks

Abstract

The recent trends of saliency modeling suggest that deep learning based saliency models are viable tools as they achieve outstanding results. However, they require large training times and expensive hardware. This study presents a prediction model of bottom up visual attention that exploit deep features from a pre-trained deep convolutional neural network. We evaluate deep features of a variety of activations from seven state-of-the-Art deep convolutional neural networks, including 35 implementations of the proposed bottom up saliency model. All implementations are evaluated over a popular dataset using three evaluation metrics. The experimental results suggest that deep features of a network trained for object classification can be used for saliency modeling. While the ranking of the saliency models is varied across scores, majority of the GoogLeNet based implementations outperformed all other implementations.