A mix-supervised unified framework for salient object detection

摘要

Recently, although deep learning network has shown its advantages in supervised salient object detection, supervised models often require massive pixel-wise annotations and learnable parameters, which seriously manacle training and testing of models. In this paper, we present a mix-supervised unified framework for salient object detection to avoid the insufficient training labels and speed training and testing up, which is composed of a region-wise stream and a pixel-wise stream. In the region-wise stream, to avoid the requirement of expensive pixel-wise annotations, an improved energy equation based manifold learning algorithm is employed, by which accurate object location and prior knowledge are introduced by the unsupervised learning. In the pixel-wise stream, to alleviate the problem of time-consuming, a simplified bi-directional reuse network is introduced, which can obtain clear object contour and competitive performance with fewer parameters. To relieve the bottleneck pressure of parallel training and testing, each steam is directly connected to its pre-processed color feature and post-processing refinement. Extensive experiments demonstrate that each component contributes to the final results and complement each other perfectly.