Shape Retention Based 3D Point Cloud Reconstruction from a Single Image

The reconstruction of 3D point clouds has always been a difficult problem, and it is a great challenge to highlight their edge characteristics for the constructed point clouds. Existing point cloud reconstruction methods ignore the problem of blurred point cloud edges. This paper introduces the shape retention point cloud reconstruction (SRPCR), which is a multi-stage generative network to enhance 3D reconstruction results from 2D images. The Fusion of the image features and edge information is utilized to generate a more accurate sparse point cloud. The high-resolution dense point clouds are generated by adding channel attention mechanisms to the point cloud generation adversarial network, and more detailed boundary information is achieved from a single image. The experiments demonstrate SRPCR outperforms existing point cloud reconstruction methods on the ShapeNet dataset.