SIHENet: Semantic Interaction and Hierarchical Embedding Network for 360° Salient Object Detection

With the application of panoramic imaging instruments in environmental perception and measurement, 360° salient object detection (SOD) has become a concern. However, how to reduce the serious distortion of 360° images is still an open question. In this article, we fully exploit the intrinsic information complementarity of cross-projection and cross-level features and innovatively propose a semantic interaction and hierarchical embedding network (SIHENet) to realize 360° SOD. In terms of fusion of different projection features, considering that the distortion of cubemap projection (CMP) images is relatively small, we propose a multiscale hierarchical embedding (MSHE) module to establish the context of CMP images and embed equirectangular projection (ERP) images in it to reduce distortion. In terms of cross-level feature fusion, in order to solve the feature dilution problem in feature transfer, we propose a multiprojection semantic interaction (MPSI) module to integrate the high-level information of ERP and CMP images. In order to solve the edge ambiguity problem, we propose a multilevel semantic matching (MLSM) module to enhance the edge information at different levels of ERP image. Extensive experiments on three public 360° datasets demonstrate the competitive performance of the proposed model in comparison to state-of-the-art (SOTA) 360° SOD models.