Consistency perception network for 360° omnidirectional salient object detection

With the gradual popularization of panoramic cameras and the rapid development of computer vision technology, the research on salient object detection (SOD) in 360 degrees omnidirectional images has attracted great attention. Different from traditional 2D images, 360 degrees omnidirectional images generally have some drawbacks such as projection distortion, complex scenes, and small objects. To address these knotty issues, in this paper, we propose a novel C onsistency P erception Net work (CPNet) for 360 degrees omnidirectional salient object detection, which can reliably localize salient regions in 360 degrees omnidirectional images. We adopt a split-splice strategy on equirectangular 360 degrees images to solve the distortion problem caused by the discontinuities of objects on the projection boundary, which restores the structure of objects and ensures the integrity of scene to the greatest extent. Inspired by the human visual perception mechanism, we deploy a bidirectional scale- aware module (BSM), which uses different convolutional dilation rates to simulate different receptive fields, performs hierarchical perceptual learning in series and conducts bidirectional guidance positioning in parallel. In addition, a consistent context-aware module (CCM) is designed to facilitate consistent learning of salient regions from different attention perspectives. It not only completely guarantees the global uniformity, but also accurately preserves the local details. Under the gradual feedback and guidance of coarse prediction maps, the edge sharpening module (ESM) can depict the details of salient objects more accurately, thereby generating high-quality saliency maps with clear boundaries. Extensive experiments on three public 360 degrees SOD datasets demonstrate that the proposed CPNet achieves comparable and competitive performance in terms of effectiveness and efficiency when compared with the cutting-edge SOD methods.