A consensus-driven approach for structure and texture aware depth map upsampling.

Area of Science:

• Computer Vision
• Image Processing
• Computer Graphics

Background:

• Depth map upsampling is crucial for 3D scene understanding.
• Existing methods often fail due to incorrect assumptions about depth and color correlations, causing artifacts.
• Accurate depth information is vital for applications like augmented reality and robotics.

Purpose of the Study:

• To develop a novel method for increasing the spatial resolution of depth maps.
• To overcome limitations of previous methods by addressing the correlation between depth discontinuities and color boundaries.
• To improve the quality and accuracy of upsampled depth maps.

Main Methods:

• A framework is proposed to divide color images into distinct regions based on depth map segmentation.
• Region-specific strategies are applied: interpolation in continuous regions and depth-histogram methods in discontinuous regions.
• A consensus of multiple super-pixel segmentation methods guides color image segmentation for robustness.

Main Results:

• The proposed method significantly improves depth map upsampling quantitatively and qualitatively.
• Artifacts common in previous approaches are reduced by region-specific processing.
• The method demonstrates robustness in handling real-world data, including occluded regions.

Conclusions:

• The developed technique effectively enhances depth map resolution by intelligently utilizing color image information.
• The region-based approach provides a more accurate and artifact-free depth map upsampling solution.
• The method is adaptable for real-world scenarios with sensor displacement challenges.