Reconstruction of 3D mesh from 2D image using deep learning
This paper evaluates the feasibility of deep learning for monocular depth estimation in the reconstruction of 3D meshes. Three deep learning models were used to generate a depth map, and three surface reconstruction algorithms were used to reconstruct the mesh. The different combinations were explor...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Final Year Project |
Language: | English |
Published: |
Nanyang Technological University
2022
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/156658 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | This paper evaluates the feasibility of deep learning for monocular depth estimation in the reconstruction of 3D meshes. Three deep learning models were used to generate a depth map, and three surface reconstruction algorithms were used to reconstruct the mesh. The different combinations were explored, and the best combination was found to be the supervised deep learning model, Dense-Depth, paired with the surface reconstruction using alpha shapes. The meshes produced were able to capture major features of the scene, but tended to have gaps within the mesh, and the depth of the surface would fluctuate. To create a higher quality mesh, the accuracy and resolution of the depth estimation models would have to be improved first. This final year project is part of research project “Artificial Intelligence for Smart Image Understanding” at Rolls-Royce@NTU Corporate Lab. |
---|