RGB-D Salient Object Detection With Ubiquitous Target Awareness

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Conventional RGB-D salient object detection relies heavily on depth data quality, which can be unreliable or unavailable.
• Existing methods struggle with depth data limitations, impacting salient object detection accuracy.

Purpose of the Study:

• To develop a novel depth-awareness framework for RGB-D salient object detection (SOD) that is depth-free during inference.
• To propose the Ubiquitous Target Awareness (UTA) network to address key challenges in RGB-D SOD.

Main Methods:

• A depth awareness module with adaptive depth-error weights to mine ambiguous regions.
• Spatial-aware cross-modal and channel-aware cross-level interactions to leverage boundary cues and salient channels.
• A gated multi-scale predictor for object saliency perception across different scales.

Main Results:

• The proposed UTA network achieves high performance and is depth-free for inference, running at 43 FPS.
• Outperforms state-of-the-art methods on five public RGB-D SOD benchmarks.
• Demonstrates extensibility on five public RGB SOD benchmarks.

Conclusions:

• The novel depth-awareness framework effectively addresses RGB-D salient object detection challenges.
• The UTA network provides accurate, real-time, and versatile salient object detection capabilities.
• This approach advances salient object detection by reducing reliance on high-quality depth data.