A Transformer-based Network for Multi-view 3D Mesh Generation

This is the authors' abstract. We don't add key points for this paper.

Three-dimensional (3D) reconstruction has been widely used in cultural relics protection, architectural design, e-commerce and other fields. In recent years, deep network-based 3D mesh generation has become mainstream. However, the existing 3D mesh generation based on convolutional neural network mainly focuses on local information, and can only update the feature of the central vertex through adjacent vertices, which greatly affects the efficiency of feature transfer between remote vertices. In order to obtain better 3D reconstruction results, we explore a 3D mesh reconstruction network combining transformer and graph convolution, which effectively takes into account the utility of global and local features. The proposed method alternately uses transformers and graph convolutions at multiple scales, so as to effectively perceive the global structure of the target of interest through the transformer and perceive local details through the graph convolution. Furthermore, our method effectively fuses information from different viewpoints to improve the performance of 3D mesh reconstruction. Experimental results show that the proposed method effectively improves the accuracy of 3D mesh reconstruction.