360$^{\circ }$ Omnidirectional Salient Object Detection with Multi-scale Interaction and Densely-Connected Prediction.

Haowei Dai,Liuxin Bao,Kunye Shen,Xiaofei Zhou,Jiyong Zhang

DOI: https://doi.org/10.1007/978-3-031-46305-1_35

2023-01-01

Abstract:Despite the fact that 360 $$^{\circ }$$ cameras have become affordable, the efforts in salient object detection (SOD) of 360 $$^{\circ }$$ omnidirectional images are still relatively few. The performance of the SOD task is significantly hindered by the severe distortion and the large range of object sizes in omnidirectional images. Therefore, we propose a Multi-scale Interaction and Densely-connected Prediction network (MIDP-Net) to counteract distortion and extract multi-scale information in 360 $$^{\circ }$$ omnidirectional images. Specifically, we introduce a Multi-Level Features Interaction (MLFI) module that firstly integrates the features with different receptive fields from adjacent levels, and then fuses the semantic cues from upper-level decoder to obtain the accurate multi-scale information. In addition, to overcome the distortion of 360 $$^{\circ }$$ omnidirectional image, we design a Two-Pass Decoder (TPD) module, which further strengthen the semantic information by secondary decoding and combines the edge cues and saliency cues to optimizes the details of prediction map. The comprehensive experiments on two public 360 $$^{\circ }$$ omnidirectional datasets indicate that the proposed MIDP-Net outperforms the state-of-the-art methods.