Improved Saliency Detection in RGB-D Images Using Two-phase Depth Estimation and Selective Deep Fusion.

Area of Science:

• Computer Vision
• Image Processing
• Artificial Intelligence

Background:

• RGB-D saliency detection relies heavily on depth information for foreground object distinction.
• Depth data quality significantly impacts saliency detection performance.
• Existing methods struggle with low-quality depth or low-contrast salient objects.

Purpose of the Study:

• To develop a method for estimating high-quality complementary depth information (Depth+) for RGB-D saliency detection.
• To improve the robustness and accuracy of saliency detection in challenging scenarios with degraded depth data.

Main Methods:

• Estimated high-quality depth (Depth+) by retrieving similar images and building inter-image nonlocal correspondences.
• Employed a depth-transferring strategy using inter-image correspondences for coarse depth estimation.
• Refined depth estimation using fine-grained, object-level correspondences and a saliency prior.
• Developed a selective deep fusion network to integrate original depth and Depth+ for optimal saliency boundary detection.

Main Results:

• The proposed Depth+ provides more informative depth data compared to the original depth.
• The selective deep fusion network effectively balances original depth and Depth+ for improved saliency detection.
• Enhanced performance in RGB-D saliency detection, particularly in scenarios with low color contrast and poor depth quality.

Conclusions:

• The novel Depth+ estimation significantly boosts RGB-D saliency detection performance.
• The proposed selective deep fusion network effectively leverages complementary depth information.
• This approach offers a promising solution for robust saliency detection with imperfect depth data.