Single Image 3D Reconstruction Based on Attention Mechanism and Graph Convolution Network

Abstract

∗ This paper innovatively proposes a channel attention mechanism and graph convolutional network model adapted to 3D reconstruction, and combines the target detection model to construct a neural network that generates a target 3D model from a single RGB image. The model first generates the target 3D Voxels, and further generates a more refined 3D Meshes model through the graph convolutional network. Compared with the control group algorithm, the Pix3D dataset AP mesh index has been improved by 2.8%, which fully proves the effectiveness of the model in single-image 3D reconstruction. The experimental results show that the algorithm has good usability in the 3D reconstruction of actual mesh points.