DF-Net: Dynamic and Folding Network for 3D Point Cloud Completion

Abstract

The development of 3D sensors encourages researchers to process point cloud data directly. Point cloud data requires less memory but conveys more detailed 3D shape information. However, because of occlusion, sensing distance and other reasons, sensors usually cannot get a complete 3D shape. In this paper, we propose a Dynamic and Folding Network (DF-Net) to address the precise point cloud completion problem. Existing completion networks generate the overall shape of a point cloud from an incomplete point cloud. In this paper, we complete the missing part based on the existing part instead. We use a dynamic graph network to better extract local features of points in the neighborhood points. A FoldBlock is used to refine the prediction of the missing part. We validate our method with two benchmarks, ShapeNet-13 and ShapeNet-55. Both qualitative and quantitative experimental results show the proposed method achieves improvement over some state-of-the-art methods. Code is available at https://github.com/yiqisetian/DF-Net.

3dcascade-Gan: Shape Completion from Single-View Depth Images

“…Xiao et al [53] proposed to use folding blocks on latent features to enhance the reconstruction for regions with missing points.…”