Robust salient object detection for RGB images

Recent advances in supervised salient object detection modeling have resulted in significant performance improvements on benchmark dataset. However, most of the existing salient object detection models assume that an image contains at least one salient object. Such an assumption that often leads to their effectiveness may be impaired once they are applied to real-world scenes. To solve the problem, salient object existence prediction designed to judge whether the image contains salient objects is introduced into deep network to learn a better salient object detection model. For dense salient object detection task, high-level semantic feature is progressively hybrid upsampled from deep to shallow to remedy the spatial information loss guided by higher layer feature and saliency existence information. Our model is aware of non-salient image that contains no salient objects at all and thus reduces the false-positive rate. Experimental results show that our model wins similar multi-task model and outperforms state-of-the-art models in robustness and accuracy.