DeepAttent: Saliency Prediction with Deep Multi-scale Residual Network

Predicting where humans look in a given scene is a well-known problem with multiple applications in consumer cameras, human-computer interaction, robotics, and gaming. With large-scale image datasets available for human fixation, it is now possible to train deep neural networks for generating a fixationmap. Human fixations are a function of both local visual features and global context. We incorporate this in a deep neural network by using global and local features of an image to predict human fixations. We sample multi-scale features of the deep residual network and introduce a new method for incorporating these multi-scale features for the end-to-end training of our network. Our model DeepAttent obtains competitive results on SALICON and iSUN datasets and outperforms state-of-the-art methods on various metrics.