- Conference Article
- 10.1109/gecost66002.2025.11324972
Deep Learning-Based 3D Reconstruction from Monocular Inputs
- Nov 12, 2025
- Ariel Starling + 5 more +5
Three-dimensional (3D) reconstruction from two-dimensional (2D) images and monocular videos remains a challenge in computer vision and graphics, particularly for complex surfaces and non-ideal imaging conditions. This short review, conducted using PRISMA guidelines, surveys deep-learning based approaches published between 2020 and 2025, emphasising neural implicit representations, real-time 3D reconstruction, transformer-based architectures and hybrid methods. Neural Implicit methods such as Neuralangelo and NeuS demonstrate accurate and complete surface modelling, with Neuralangelo more notably achieving higher fidelity standards than its predecessor NeuS while being 7% more efficient. For real-time applications, RTG-SLAM offers efficient, coherent scene reconstruction through recent adaptive 3D Gaussian Splatting methods. Transformer-based frameworks like EVolT address persistent multi-view attention challenges, though they remain restricted by voxel resolution and memory demands. These emerging technologies have broad implications across domains including AR, cultural heritage, medical imaging, object prototyping, and plant phenotyping.
Read more