This paper presents a novel approach to visual saliency modeling using multiscale deep features extracted from convolutional neural networks (CNNs). The authors introduce a neural network architecture that combines fully connected layers on top of CNNs to extract features at three different scales. They propose a refinement method to enhance the spatial coherence of saliency results and demonstrate that aggregating multiple saliency maps computed for different levels of image segmentation can further improve performance. To promote research and evaluation, they construct a large dataset of 4447 challenging images with pixel-wise saliency annotations. Experimental results show that their method achieves state-of-the-art performance on public benchmarks, improving the F-Measure by 5.0% and 13.2% on the MSRA-B dataset and their new dataset (HKU-IS), and reducing the mean absolute error by 5.7% and 35.1% on these datasets. The paper also discusses the effectiveness of the proposed features, spatial coherence, and multilevel decomposition in the saliency model.This paper presents a novel approach to visual saliency modeling using multiscale deep features extracted from convolutional neural networks (CNNs). The authors introduce a neural network architecture that combines fully connected layers on top of CNNs to extract features at three different scales. They propose a refinement method to enhance the spatial coherence of saliency results and demonstrate that aggregating multiple saliency maps computed for different levels of image segmentation can further improve performance. To promote research and evaluation, they construct a large dataset of 4447 challenging images with pixel-wise saliency annotations. Experimental results show that their method achieves state-of-the-art performance on public benchmarks, improving the F-Measure by 5.0% and 13.2% on the MSRA-B dataset and their new dataset (HKU-IS), and reducing the mean absolute error by 5.7% and 35.1% on these datasets. The paper also discusses the effectiveness of the proposed features, spatial coherence, and multilevel decomposition in the saliency model.