HiDAnet: RGB-D Salient Object Detection via Hierarchical Depth Awareness.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• RGB-D saliency detection fuses color and depth data to identify important image regions.
• Current methods struggle to integrate fine-grained details with semantic cues, limiting performance with similar-looking objects at different depths.

Purpose of the Study:

• To propose a novel Hierarchical Depth Awareness network (HiDAnet) for improved RGB-D saliency detection.
• To enhance the fusion of multi-modal and multi-level features by leveraging geometric priors.

Main Methods:

• Developed a granularity-based attention scheme to enhance RGB and depth feature discrimination.
• Introduced a unified cross dual-attention module for coarse-to-fine multi-modal and multi-level feature fusion.
• Utilized a shared decoder and multi-scale loss for hierarchical information aggregation.

Main Results:

• HiDAnet demonstrated superior performance on challenging benchmark datasets.
• The proposed method significantly improved upon state-of-the-art RGB-D saliency detection techniques.
• Achieved favorable results by effectively integrating hierarchical depth information.

Conclusions:

• The Hierarchical Depth Awareness network (HiDAnet) offers an effective approach for RGB-D saliency detection.
• The proposed multi-modal and multi-level fusion strategy addresses limitations in existing methods.
• HiDAnet shows strong potential for applications requiring precise object localization based on visual and depth cues.