ICNet: Information Conversion Network for RGB-D Based Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• RGB-D based salient object detection (SOD) methods utilize depth maps for enhanced performance.
• Existing fusion strategies in RGB-D SOD often fail to fully capture complex correlations and cross-modal information.
• Current methods treat information from different sources without discrimination, limiting potential.

Purpose of the Study:

• To propose a novel Information Conversion Network (ICNet) for improved RGB-D based SOD.
• To address limitations in existing fusion strategies by exploring cross-modal complementarity and continuity.
• To develop methods for interactively and adaptively fusing high-level RGB and depth features.

Main Methods:

• Developed an Information Conversion Network (ICNet) using a siamese structure with an encoder-decoder architecture.
• Introduced an Information Conversion Module (ICM) for interactive and adaptive fusion of RGB and depth features.
• Designed a Cross-modal Depth-weighted Combination (CDC) block to discriminate and enhance cross-modal features.

Main Results:

• ICNet demonstrated superior performance compared to 15 state-of-the-art RGB-D SOD methods.
• Experiments on five datasets validated the effectiveness of the proposed ICM and CDC block.
• The network successfully enhanced RGB features with depth features at each level.

Conclusions:

• The proposed ICNet effectively addresses limitations in current RGB-D SOD fusion strategies.
• The ICM and CDC block significantly contribute to improved salient object detection performance.
• ICNet offers a promising advancement in leveraging RGB-D data for accurate SOD.