Photometric Depth Super-Resolution.

Area of Science:

• Computer Vision
• Photogrammetry
• Machine Learning

Background:

• RGB-D sensors provide depth information but often at lower resolution than companion RGB images.
• Upsampling depth maps to match RGB image resolution is crucial for many applications.
• Existing photometric methods often rely on specific reflectance assumptions.

Purpose of the Study:

• To explore photometric techniques for upsampling low-resolution depth maps from RGB-D sensors.
• To develop methods that relax constraints on target reflectance properties.
• To evaluate the effectiveness of proposed depth map enhancement techniques.

Main Methods:

• A single-shot variational approach for depth map upsampling.
• Utilizing deep neural networks for reflectance estimation in specific object classes (e.g., faces).
• A multi-shot strategy with varying lighting conditions, requiring no prior reflectance knowledge.

Main Results:

• The single-shot method is effective for piecewise-constant reflectance targets.
• Deep learning relaxes reflectance model dependency for specific object classes.
• The multi-shot approach offers training-free enhancement but needs specialized hardware.
• Both methods demonstrate effectiveness on synthetic and real-world data.

Conclusions:

• Photometric techniques, enhanced by deep learning or multi-shot strategies, can effectively upsample low-resolution depth maps.
• The choice of method depends on reflectance properties, object class, and available acquisition setup.
• These advancements improve the integration of depth and color information in computer vision tasks.