Structure Selective Depth Superresolution for RGB-D Cameras.

Area of Science:

• Computer Vision
• Image Processing
• Computational Photography

Background:

• Standard depth upsampling methods (joint filtering, quadratic optimization) produce artifacts like texture copying and depth bleeding.
• These artifacts stem from structural discrepancies between depth and intensity data from different sensors.
• Existing methods struggle with the differing distributions and formations of depth and intensity discontinuities.

Purpose of the Study:

• To develop a high-quality depth superresolution technique.
• To address and mitigate texture copying and depth bleeding artifacts in image-guided depth upsampling.
• To improve the accuracy and fidelity of depth maps generated from multiple sensor data.

Main Methods:

• Formulation of an optimization model incorporating a nonconvex regularizer.
• Utilizing nonlocal affinity in a high-dimensional feature space for precise depth boundary localization.
• Iterative handling of structural differences between depth and intensity images.
• Development of a fast alternating direction method of multipliers (ADMM) algorithm for optimization.

Main Results:

• Significant reduction in texture copying and depth bleeding artifacts across diverse datasets.
• Demonstrated superiority over existing depth superresolution methods on synthetic and real-world data.
• The proposed ADMM solver offers a noticeable speed-up compared to traditional majorize-minimize algorithms.

Conclusions:

• The proposed nonconvex regularizer and nonlocal affinity approach effectively resolves structural discrepancies between sensor data.
• This method achieves high-quality depth superresolution with reduced artifacts.
• The efficient ADMM solver makes the technique practical for real-world applications.