CIA-Net: Cross-Modal Interaction and Depth Quality-Aware Network for RGB-D Salient Object Detection

Depth information has been proven beneficial in RGB-D salient object detection (SOD). However, the depth maps are usually of low quality in existing RGB-D SOD datasets. Most RGB-D SOD models lack cross-modal interaction or fail to consider the quality of depth maps during cross-modal interaction, which could lead to inaccurate encoder features when facing low-quality depth maps. In this paper, we propose a novel network called CIA-Net to measure depth map quality and effectively integrate complementary information across modalities. Through the depth quality-aware module, feature alignment of low-order information generates weights to represent the quality of the depth map. The weighted modality interaction module controls the weight of the depth map to perform interactions, which applies attention mechanisms to interact with the cross-modal features at each scale. Extensive experiments show that our proposed model significantly outperforms ten existing state-of-the-art models on four challenging benchmark datasets. Codes and results will be available after this work is accepted.