Salient Object Detection From Unlabeled Images

Abstract Salient object detection via deep neural networks usually needs a large amount of images with human annotation. To avoid laborious and consuming annotation, we propose robust unsupervised salient method three stages in this work. Our first uses unlabeled data to generate an activation map which indicates the coarse location object. Then one scribble generation based on is proposed, provides foreground background high confidence. Finally, model trained supervision generated scribble. Performance comparison carried four public datasets, showing that our significantly outperforms state-of-the-art methods, also achieves better performance than some weakly supervised methods.