Dynamic Selective Network for RGB-D Salient Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Image Processing

Background:

• RGB-D saliency detection integrates RGB images and depth maps.
• Existing methods often fail to address modality differences, impacting performance.
• Challenging scenes require robust fusion strategies for accurate salient object detection.

Purpose of the Study:

• To propose a novel RGB-D saliency detection model, Dynamic Selective Network (DSNet).
• To effectively utilize the complementarity between RGB and depth data.
• To enhance salient object detection (SOD) performance in complex scenarios.

Main Methods:

• A cross-modal global context module (CGCM) for high-level semantic information acquisition.
• A dynamic selective module (DSM) for mining cross-modal complementary information.
• Gated and pooling-based selection for multi-level, multi-scale information optimization and boundary refinement.

Main Results:

• DSNet demonstrates superior performance compared to 17 state-of-the-art RGB-D SOD models.
• The model achieves competitive and excellent results across eight public RGB-D datasets.
• Effective integration of RGB and depth information leads to high-quality saliency maps.

Conclusions:

• DSNet successfully addresses the limitations of existing RGB-D saliency detection methods.
• The proposed dynamic selective approach enhances the utilization of cross-modal information.
• DSNet provides a robust solution for salient object detection in diverse and challenging RGB-D scenes.