Journal Article

Enhanced multi view 3D reconstruction with improved MVSNet

Wang, Zhenfei, Zhu, Zhenfang, Wang, Peng, Fu, Chen, Zhang, Guangyuan, Li, Guangchen, Li, Kefeng

2024

Overview

Although 3D reconstruction has been widely used in many fields as a key component of environment perception, existing technologies still have the potential for further improvement in 3D scene reconstruction. We propose an improved reconstruction algorithm based on the MVSNet network architecture. To glean richer pixel details from images, we suggest deploying a DE module integrated with a residual framework, which supplants the prevailing feature extraction mechanism. The DE module uses ECA-Net and dilated convolution to expand the receptive field range, performing feature splicing and fusion through the residual structure to retain the global information of the original image. Moreover, harnessing attention mechanisms refines the 3D cost volume's regularization process, bolstering the integration of information across multi-scale feature volumes, consequently enhancing depth estimation precision. When assessed our model using the DTU dataset, findings highlight the network's 3D reconstruction scoring a completeness (comp) of 0.411 mm and an overall quality of 0.418 mm. This performance is higher than that of traditional methods and other deep learning-based methods. Additionally, the visual representation of the point cloud model exhibits marked advancements. Trials on the Blended MVS dataset signify that our network exhibits commendable generalization prowess.

• Publisher: Nature Publishing Group UK,Nature Publishing Group,Nature Portfolio
• Subject: 639/166 / 639/705/117 / Accuracy / Algorithms / Deep learning / Design / Humanities and Social Sciences / multidisciplinary / Neural networks / Receptive field / Science / Science (multidisciplinary)