Unsupervised Salient Object Detection Based on Depth-Induced Pseudo-Labels Updating

The performance of RGB-D salient object detection have improved prominently in recent years, owing to the rapid development of deep learning algorithms. However, recent state-of-the-art RGB-D salient object detection are mainly based on the supervised learning which demand large quantity of ground truth being thoroughly annotated at pixel-level. They are really laborious and time-consuming. Furthermore, the RGB-D salient object detection training with supervised learning are not suitable for various practical scenarios. In this paper, we tackle the challenging task of unsupervised salient object detection by leveraging the pseudo-labels automatic updating and the depth-induced uncertainty module, named RGB-D USOD Based on Depth-induced Pseudo-labels Updating (DPU). There are three main steps for our proposed methods. The first and second steps are used to generate and optimize the pseudo-labels based on center-dark channel prior and VGG backbone. The first step relies on handcrafted feature representations while the second one depends on semantic representations. Last but not the least, the DPU is proposed to update the pseudo ground truth during the training process and predict the final salient object. Extensive experiments demonstrate the superior efficiency and effectiveness of our approach in tackling the challenging unsupervised RGB-D salient object detection, compared with 16 unsupervised salient object detectors.