A Survey of Depth Completion Techniques from Classical Approaches to Deep Learning Models
Harju, Teemu (2025)
2025
Tieto- ja sähkötekniikan kandidaattiohjelma - Bachelor's Programme in Computing and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2025-12-22
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2025122012027
https://urn.fi/URN:NBN:fi:tuni-2025122012027
Tiivistelmä
Depth completion is a fundamental task in computer vision that addresses the challenge of converting sparse, incomplete depth measurements from sensors like LiDAR into dense and accurate depth maps. This capability is crucial for a wide range of applications, including autonomous driving, robotics, and augmented reality, where a precise understanding of the 3D environment is essential.
This thesis provides a comprehensive survey of depth completion techniques, tracing their evolution from classical methods to modern deep learning models. We begin by examining traditional approaches, such as interpolation, filtering, and optimization-based algorithms. These methods are computationally efficient but often struggle with complex scenes, leading to over-smoothed results and blurred object boundaries.
In contrast, deep learning has significantly advanced the field by leveraging data-driven models capable of learning semantic and geometric cues directly from large datasets. This thesis examines key deep learning paradigms such as multi-modal fusion of RGB and depth data, transformer-based architectures for capturing global context, diverse loss functions for structured depth prediction, and spatial propagation networks for iterative refinement. Furthermore, it discusses emerging trends like self-supervised and unsupervised depth completion, hybrid models combining classical priors with neural networks, and lightweight architectures designed for real-time deployment.
The thesis concludes by analyzing the trade-offs between traditional and learning-based methods, identifying persistent challenges such as extreme input sparsity, domain generalization, and fine-grained boundary preservation, and highlighting promising directions for future research. By consolidating a wide body of literature, this work serves as an accessible reference for understanding the progression of depth completion techniques and provides insights into ongoing developments toward robust, accurate, and efficient depth perception systems.
This thesis provides a comprehensive survey of depth completion techniques, tracing their evolution from classical methods to modern deep learning models. We begin by examining traditional approaches, such as interpolation, filtering, and optimization-based algorithms. These methods are computationally efficient but often struggle with complex scenes, leading to over-smoothed results and blurred object boundaries.
In contrast, deep learning has significantly advanced the field by leveraging data-driven models capable of learning semantic and geometric cues directly from large datasets. This thesis examines key deep learning paradigms such as multi-modal fusion of RGB and depth data, transformer-based architectures for capturing global context, diverse loss functions for structured depth prediction, and spatial propagation networks for iterative refinement. Furthermore, it discusses emerging trends like self-supervised and unsupervised depth completion, hybrid models combining classical priors with neural networks, and lightweight architectures designed for real-time deployment.
The thesis concludes by analyzing the trade-offs between traditional and learning-based methods, identifying persistent challenges such as extreme input sparsity, domain generalization, and fine-grained boundary preservation, and highlighting promising directions for future research. By consolidating a wide body of literature, this work serves as an accessible reference for understanding the progression of depth completion techniques and provides insights into ongoing developments toward robust, accurate, and efficient depth perception systems.
Kokoelmat
- Kandidaatintutkielmat [10837]