G2IFu: Graph-based implicit function for single-view 3D reconstruction

Rongshan Chen, Yuancheng Yang, Chao Tong

Abstract: As the demand for 3D models increases, there is growing interest in reconstructing 3D objects from images using AI. In this paper, we propose G2IFu, a graph-based implicit function that successfully reconstructs a highly detailed 3D object mesh from a single image. Unlike previous methods that learn implicit functions with points, G2IFu aims to map graphs to implicit values. We make the following contributions: (1) Compared to independent 3D points, graphs have a larger perception space and contain specific spatial structure information. Therefore, we extend a 3D point p<math><mi is="true">p</mi></math> to a graph Gp<math><msub is="true"><mrow is="true"><mi mathvariant="script" is="true">G</mi></mrow><mrow is="true"><mi is="true">p</mi></mrow></msub></math> by generating hypothesis points and establishing edges between them. We then predict the corresponding implicit value using a graph convolution network. Our experiments show that this method can effectively improve the prediction accuracy of implicit functions. (2) We introduce a prior boundary loss based on Gp<math><msub is="true"><mrow is="true"><mi mathvariant="script" is="true">G</mi></mrow><mrow is="true"><mi is="true">p</mi></mrow></msub></math> to make the network pay more attention to the “key” points near the shape surface. To the best of our knowledge, G2IFu is the first model that introduces a graph into neural implicit representation. (3) Inspired by previous methods, we utilize the image’s global and local features to initialize Gp<math><msub is="true"><mrow is="true"><mi mathvariant="script" is="true">G</mi></mrow><mrow is="true"><mi is="true">p</mi></mrow></msub></math>. We also introduce a self-attention module into G2IFu for better performance. We conduct experiments on the ShapeNet dataset and demonstrate that G2IFu can generate higher-quality 3D object shapes than previous single-view reconstruction methods. Additionally, we extend G2IFu to multi-view 3D reconstruction and achieve good performance.