Hierarchical Feature Integration Network for RGB-D Saliency Detection

Despite their promising results, existing two-stream RGB-D saliency detection methods fall short in fully exploiting cross-modal complementary features. In this paper, we develop a novel hierarchical feature integration network (HFINet), designed to explicitly and effectively leverage the complementary nature of two-stream features while integrating multi-level spatial characteristics. Specifically, for low-level features where RGB contains richer detail than depth, we introduce a pyramid spatial fusion module (PSFM) to extract only RGB-based detail information, enhancing both detail preservation and contextual transmission. For high-level features, a pyramid feature fusion module (PFFM) is proposed to capture semantic content from RGB while aggregating contextual fusion information. Moreover, a feature interaction module (FIM) is designed to leverage depth cues to assist RGB representation, enabling accurate mining of semantic complementarity between the two modalities. Finally, a lightweight feature fusion decoder is adopted to facilitate efficient feature transformation from the encoder to the decoder. Extensive experiments on several datasets show that HFINet achieves competitive performance compared to 11 representative methods.