Three-Dimensional Reconstruction from a Single RGB Image Using Deep Learning: A Review.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Graphics

Background:

• Single-view 3D reconstruction is an ill-posed problem due to infinite possible 3D shapes projecting to one 2D image.
• Complex deformations and lack of texture further challenge accurate 3D shape recovery.
• Recent advancements in deep learning have significantly improved performance in this area.

Purpose of the Study:

• To review recent literature on single-view 3D reconstruction methods.
• To focus on deep learning approaches published between 2018 and 2021.
• To categorize and discuss various 3D shape representations and evaluation metrics.

Main Methods:

• Literature review of deep learning-based 3D reconstruction techniques.
• Categorization of methods based on output representations: depth maps, surface normals, point clouds, and meshes.
• Analysis of diverse loss functions and evaluation metrics used in recent studies.

Main Results:

• Deep learning methods show significant performance improvements for single-view 3D reconstruction.
• Various 3D representations (depth maps, point clouds, meshes) are employed, each with strengths and weaknesses.
• Lack of standardized datasets and comparison metrics hinders direct method evaluation.

Conclusions:

• Deep learning has revolutionized single-view 3D reconstruction, addressing the ill-posed nature of the problem.
• Future research may benefit from standardized datasets and evaluation protocols for better comparison.
• The reviewed methods offer diverse strategies for recovering 3D shapes from limited 2D information.