Deep Depth Completion From Extremely Sparse Data: A Survey.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Depth completion predicts dense pixel-wise depth from sparse sensor data (e.g., LiDAR).
• Crucial for applications like autonomous driving, 3D reconstruction, augmented reality, and robot navigation.
• Deep learning methods currently dominate depth completion research.

Purpose of the Study:

• To provide a comprehensive literature review of depth completion.
• To help readers understand current research trends and advances.
• To propose a novel taxonomy for categorizing existing methods.

Main Methods:

• Investigated network architectures, loss functions, datasets, and learning strategies.
• Categorized existing depth completion methods using a novel taxonomy.
• Performed quantitative performance comparisons on indoor and outdoor benchmarks.

Main Results:

• Identified key design aspects and learning strategies in deep learning-based depth completion.
• Provided a structured overview of the field through a new categorization system.
• Quantitatively compared model performance across multiple benchmark datasets.

Conclusions:

• Deep learning significantly advances depth completion capabilities.
• A comprehensive understanding of current methods and challenges is crucial.
• Future research directions are identified to address existing limitations.