Residual Geometric Feature Transform Network for 3D Surface Super-Resolution

In 3D reconstruction, how to recover high-resolution 3D surface details from the existing low-resolution 3D surface is still a challenging problem. Due to the unstructured and irregular characteristics of 3D data, it is usually difficult to obtain extremely dense 3D surface and capture detailed local features. To tackle this problem, this article introduces an effective deep convolutional network, namely RGFTNet, to perform 3D surface super-resolution in 2D normal domain. To restore dense surface details and learn sharp geometry structures simultaneously, a shape prior acquisition method is designed to achieve the high-quality shape normal from the input low-resolution one. Subsequently, the extracted shape normal as the shape prior is incorporated into a deep convolutional network through the Geometric Feature Transform (GFT) layer. Experimental results show the superiority of the proposed RGFTNet over several recent advances on both the computer-generated and the real-world data.