A Comprehensive Survey of Depth Completion Approaches

Area of Science:

• Computer Vision
• Robotics
• 3D Sensing

Background:

• LiDAR sensors produce sparse and noisy depth maps, hindering accurate 3D scene reconstruction.
• Depth completion aims to generate dense depth maps from sparse inputs, essential for autonomous systems.
• Modern methods leverage RGB images for guidance, improving upon earlier unguided techniques.

Purpose of the Study:

• To provide a comprehensive survey of existing depth completion methods.
• To introduce a novel taxonomy for classifying depth completion approaches.
• To review state-of-the-art techniques for LiDAR depth map enhancement.

Main Methods:

• Categorization of depth completion methods into unguided and image-guided approaches.
• Subdivision of image-guided methods into multi-branch and spatial propagation networks.
• Detailed review of techniques within each category, including image-guided filtering.

Main Results:

• A structured overview of the depth completion literature is presented.
• Quantitative results are provided for various methods on benchmark datasets (KITTI, NYUv2).
• The survey facilitates understanding of current advancements in LiDAR depth completion.

Conclusions:

• The paper offers a foundational taxonomy for depth completion research.
• It highlights the effectiveness of image-guided methods for enhancing sparse LiDAR data.
• This work serves as a valuable resource for researchers in autonomous driving and robotics.