Learning Selective Mutual Attention and Contrast for RGB-D Saliency Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Effective cross-modal fusion of RGB and depth data is crucial for RGB-D salient object detection.
• Existing fusion methods (early, result, feature) suffer from distribution gaps, information loss, or limited low-order fusion.
• Previous approaches often fail to fully leverage complementary information between RGB and depth modalities.

Purpose of the Study:

• To propose a novel mutual attention model for enhanced RGB-D salient object detection.
• To address the limitations of existing fusion strategies by enabling high-order cross-modal interaction.
• To improve the robustness and performance of salient object detection models, especially with potentially low-quality depth data.

Main Methods:

• Developed a mutual attention model that fuses attention and context from different modalities.
• Utilized non-local attention for long-range contextual dependency propagation between RGB and depth streams.
• Incorporated contrast inference and selective attention to reweight depth cues and create a unified model.

Main Results:

• The proposed mutual attention model demonstrated significant effectiveness in RGB-D salient object detection.
• The model achieved high-order and trilinear cross-modal interaction, overcoming limitations of point-to-point fusion.
• A new, large-scale, high-quality RGB-D salient object detection dataset was constructed to facilitate model training and evaluation.

Conclusions:

• The proposed mutual attention mechanism offers a superior approach for fusing cross-modal information in RGB-D salient object detection.
• The selective attention module effectively handles potentially noisy depth data, improving model reliability.
• The new dataset will advance research and development in the field of RGB-D salient object detection.