Fusion Multi-scale Features and Attention Mechanism for Multi-view 3D Reconstruction

In order to solve the problem that it is difficult to carry out the reconstruction accurately and completely in the process of multi-view 3D reconstruction, we puts forward to a 3D reconstruction method that combined with multi-scale features and attention mechanisms. To begin with, multi-scale features are extracted from the input source and reference images using a Feature Pyramid Network that adds a bottom-up structure to obtain the connections between the different scales of the feature map. What’s more, the attention mechanism is used to optimize the cost volume regularization process and improve the accuracy of depth estimation. In the end, the improved network is trained. About the experimental results of DTU dataset, it can be seen that the 3D reconstruction results of this network reached 0.358mm and 0.366mm in Completeness (Comp) and Overall respectively, which are higher than traditional methods as well as some 3D reconstruction methods that are based on deep learning methods. At the same time, the results of visualization in the point cloud model have been significantly improved. In addition, experimental results on Tanks and Temples dataset show that this network also have a better generalization performance.