EF-Net: RGB-D-based Saliency Detection Using Information Extraction and Fusion Network

RGB-D-based salient object detection (SOD) methods are increasingly popular, which import depth information to improve SOD performance. The emergence of depth maps raises the question of how to fuse RGB images and depth maps appropriately? The existing fusion strategies can be roughly divided into three categories: early fusion, multi-scale fusion, and late fusion. However, these strategies cannot fully excavate the complex correlation between RGB images and depth maps and fail to effectively integrate multi-level features. This paper presents a novel information extraction and fusion network (EFNet) for RGB-D based SOD by employing a Siamese network with an encoder-decoder structure. Extensive experiments on 6 widely acknowledged benchmark datasets demonstrate the superiority of the proposed EF-Net over 15 state-of-the-art RGBD-based SOD methods.