Abstract: Nowadays the visual saliency prediction has become a fundamental problem in 3D imaging area. In this paper, we proposed a saliency prediction model from the perspective of addressing three aspects of challenges. First, to adequately extract features of RGB and depth information, we designed an asymmetric encoder structure on the base of U-shape architecture. Second, to prevent the semantic information between salient objects and corresponding contexts from diluting in cross-modal distillation stream, we devised a global guidance module to capture high-level feature maps and deliver them into feature maps in shallower layers. Third, to locate and emphasize salient objects, we introduced a channel-wise attention model. Finally we built the refinement stream with integrated fusion strategy, gradually refining the saliency maps from coarse to fine-grained. Experiments on two widely-used datasets demonstrate the effectiveness of the proposed architecture, and the results show that our model outperforms six selective state-of-the-art models.
Authors: Xinyue Zhang and Ting Jin (Hainan University, China)
Email: email@example.com, firstname.lastname@example.org