Title:Multi-view 3D Reconstruction based on Context Information Fusion and Full Scale Connection
Volume: 18
Author(s): Yunyan Wang, Yuhao Luo*Chao Xiong
Affiliation:
- School of Electrical and Electronic Engineering, Hubei University of Technology, Wuhan, 430068, China
Keywords:
Deep learning, three-dimensional reconstruction, multi-view stereo, full scale connection, information fusion, fringe projection profilometry.
Abstract:
Background: Multi-view stereo matching is the reconstruction of a three-dimensional
point cloud model from multiple views. Although the learn-based method achieves excellent results
compared with the traditional method, the existing multi-view stereo matching method will
lose the underlying details when extracting features due to the deepening of the number of convolutional
layers, which will affect the quality of subsequent reconstruction.
Objective: The objective of this approach is to improve the integrity and accuracy of 3D reconstruction,
and obtain a 3D point cloud model with richer texture and more complete structure.
Methods: Firstly, a context-semantic information fusion module is constructed in the feature extraction
network FPN, and the feature maps containing rich context information can be obtained
by using multi-scale dense connections.Subsequently, a full-scale jump connection is introduced
in the regularization process to capture the shallow level of detail information and deep level of
semantic information at the full scale, and capture the texture features of the scene more accurately,
so as to carry out reliable depth estimation.
Results: The experimental results on DTU dataset show that the proposed CU-MVSNet reduces
the completeness error by 3.58%, the accuracy error by 3.7%, and the overall error by 3.51%
compared with the benchmark network. It also shows good generalization on TnT dataset.
Conclusion: The CU-MVSNet method proposed in this paper can improve the completeness and
accuracy of 3D reconstruction, and obtain a 3D point cloud model with more detailed texture and
more complete structure.
