We found a match
Your institution may have access to this item. Find your institution then sign in to continue.
- Title
Multi-view 3D Reconstruction Based on Deformable Convolution and Laplace Pyramid Residuals.
- Authors
Zhaoming Hao; Ziyang Zhang; Hongyan Li; Baoqing Xu; Xiaoqiong Zhang; Meng Xu; Weifeng Wang
- Abstract
The current deep learning-based multi-view stereo point cloud reconstruction method has been found to have low reconstruction accuracy for target boundary contours. This paper proposes a high-precision and high-completeness multi-view stereo reconstruction network (DL-PatchMatchNet) based on an improved PatchMatchNet. Firstly, to increase the robustness of the model's feature extraction, a deformable convolution-based feature extraction network is proposed. Secondly, to improve model reconstruction of target contours and boundaries, the Laplace pyramid residuals are introduced to guide the decoding process of the model. Lastly, a fused loss function (GSS) is proposed to enhance the accuracy of point cloud reconstruction by simultaneously considering geometric consistency loss, structural similarity metric and smoothing loss. The results of the experimental analysis on the DTU dataset demonstrate that the DL-PatchMatchNet model exhibits a lower mean absolute error (MAE) and error rate (ER) than other competing networks. This performance is reflected in the high accuracy and completeness of reconstruction achieved by the DL-PatchMatchNet model.
- Subjects
FEATURE extraction; POINT cloud; PYRAMIDS; ERROR rates; SMOOTHING (Numerical analysis)
- Publication
IAENG International Journal of Computer Science, 2024, Vol 51, Issue 7, p896
- ISSN
1819-656X
- Publication type
Article