Webly-Supervised Learning for Salient Object Detection

•We propose a novel webly-supervised learning method for salient object detection, which requires no pixel-wise annotations.•We introduce a novel quality evaluation method that can help to pick out images with high-quality masks for training.•We present a self-training approach to boost the performance of our network by selecting more hard web images for training.•Our method outperforms existing unsupervised methods, weakly supervised methods, and even fully-supervised approaches.