Fully Convolutional DenseNet for Saliency-Map Prediction

In this paper, we propose a fully convolutional DenseNet model for saliency-map prediction (DenseSal). While the most state-of-the-art models for predicting saliency maps use shallow networks such as VGG-16, our model uses densely connected convolutional networks (DenseNet) with over 150 layers. Since DenseNet has shown the excellent results on image classification tasks, the coarse feature maps from the fully convolutional neural networks based on DenseNets were concatenated to predict saliency maps through a readout network. It is shown that the DenseNet is useful for the saliency-map prediction and achieved the state-of-the-art accuracy on the major fixation datasets.