Recent Advances in Conventional and Deep Learning-Based Depth Completion: A Survey.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Sensing

Background:

• Depth completion is crucial for applications like autonomous driving and 3D reconstruction.
• Incomplete and noisy depth maps from sensors (LiDAR, RGB-D cameras) pose significant challenges.
• Traditional methods struggle with sparse and boundary-noisy data, motivating advanced solutions.

Purpose of the Study:

• To systematically review and summarize existing research on depth completion.
• To focus on deep learning-based methods, particularly those using multiple inputs like LiDAR and RGB images.
• To identify current trends and future research prospects in the field.

Main Methods:

• Review of conventional image processing and deep learning techniques for depth completion.
• Analysis of input modalities, data fusion strategies, and loss functions.
• Emphasis on deep learning-based methods for LiDAR-image depth completion.

Main Results:

• Deep learning methods have achieved inspiring results, especially for challenging LiDAR-image depth completion.
• Systematic categorization of depth completion works based on key technical aspects.
• Identification of effective strategies for handling sparse and noisy depth data.

Conclusions:

• Depth completion is a rapidly advancing field driven by deep learning.
• Future research should explore novel data fusion and deep learning architectures.
• Continued advancements are expected to enhance performance in real-world applications.