SalDA: DeepConvNet Greets Attention for Visual Saliency Prediction