Edge-Guided Single Depth Image Super Resolution.

Area of Science:

• Computer Vision
• Image Processing
• 3D Sensing

Background:

• Consumer depth cameras offer affordable 3D data but suffer from low resolution and depth map quality issues.
• These limitations hinder the performance of depth sensing in various applications.
• Existing super-resolution methods often struggle with artifacts and edge preservation.

Purpose of the Study:

• To propose a novel framework for single depth image super-resolution.
• To enhance the resolution and quality of depth maps from consumer-grade cameras.
• To address artifacts and preserve sharp edges in upscaled depth images.

Main Methods:

• A novel framework utilizing a high-resolution edge map to guide depth image upscaling.
• Construction of the edge map via Markov random field optimization and patch synthesis.
• Exploration of patch self-similarity for edge construction with limited training data.
• Upsampling the depth image using a modified joint bilateral filter guided by the edge map.

Main Results:

• The proposed edge-based guidance effectively avoids artifacts from direct texture prediction.
• The method significantly reduces jagged artifacts and preserves sharp edges in the super-resolved depth maps.
• Experimental results show superior qualitative and quantitative performance compared to state-of-the-art methods.

Conclusions:

• The developed framework successfully enhances single depth images.
• The edge-guided approach offers a robust solution for improving depth map quality from affordable sensors.
• This method has the potential to broaden the applicability of consumer depth cameras.