Absolute and Relative Depth-Induced Network for RGB-D Salient Object Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Salient object detection in complex scenes using RGB-D data is challenging.
• Existing methods often overlook relative depth information, focusing primarily on absolute depth values.
• Effective fusion of RGB and depth features is crucial for accurate saliency detection.

Purpose of the Study:

• To propose a novel Depth-Induced Network (DIN) for RGB-D salient object detection.
• To leverage both absolute and relative depth information for enhanced feature fusion.
• To develop a lightweight yet effective model for accurate salient object localization.

Main Methods:

• Introduced an Absolute Depth-Induced Module (ADIM) for hierarchical integration of absolute depth and RGB features during encoding.
• Designed a Relative Depth-Induced Module (RDIM) to capture detailed saliency cues using contrastive and structural information from relative depth during decoding.
• Enforced in-depth fusion of RGB-D cross-modalities through the combined ADIM and RDIM.

Main Results:

• The proposed DIN accurately locates salient objects with clear boundaries, even in complex scenes.
• The network demonstrates superior performance compared to existing RGB-D salient object detection models across six challenging benchmarks.
• The DIN is a lightweight network with a significantly smaller model size than state-of-the-art algorithms.

Conclusions:

• The DIN effectively utilizes both absolute and relative depth information for robust salient object detection.
• The proposed approach offers a significant advancement in RGB-D salient object detection, particularly for complex scenarios.
• The lightweight nature of the DIN makes it a practical solution for real-world applications.