Weakly Supervised Learning for Salient Object Detection

Recent advances of supervised salient object detection models demonstrate significant performance on benchmark datasets. Training such models, however, requires expensive pixel-wise annotations of salient objects. Moreover, many existing salient object detection models assume that at least a salient object exists in the input image. Such an impractical assumption leads to less appealing saliency maps on the background images, which contain no salient objects at all. To avoid expensive strong saliency annotations, in this paper, we study weakly supervised learning approaches for salient object detection. In specific, given a set of background images and/or salient object images, where we only have annotations of salient object existence, we propose two approaches to train salient object detection models. In the first approach, we train a one-class SVM based on background superpixels. The further a superpixel is from the decision boundary of the one-class SVM, the more salient it is. The most interesting property of this approach is that we can effortlessly synthesize a set of background images to train the model. In the second approach, we present a solution toward jointly addressing salient object existence and detection tasks. We formulate salient object detection as an image labeling problem, where saliency labels of superpixels are modeled as hidden variables in the latent structural SVM framework. Experimental results on benchmark datasets validate the effectiveness of our proposed approaches.