Consistent Features from Varied Contexts and Ordered Multiple Attention Forms for RGB-D Salient Object Detection

Attention-based fusion and sophisticated multi-modal feature extraction are the key components of many state-of-the-art RGB-D salient object detection (SOD) methods. In this regard, different forms of attention providing a variety of emphasis during fusion in an appropriate order, and extraction of consistent features at multiple levels from a variety of contexts can be leveraged to improve the SOD performance further. To this end, this paper proposes an RGB-D SOD technique that learns multi-level features considering a variety of local and global contexts along with long-range dependencies. The learning also involves consistency evaluation of saliency features from pixel to region level using cascaded capsule networks to facilitate faithful preservation of intricate salient object details. The fusion of multi-modal features leading to the SOD is performed using cross-modal interactions through multiple attention forms ordered in a favorable way. Quantitative and qualitative experimental results using multiple standard datasets and measures demonstrate the effectiveness of the proposed SOD framework in comparison with the state-of-the-art. An ablation study establishes the utility of the various components of the proposed approach.