PUI-Net: A Point Cloud Upsampling and Inpainting Network

Upsampling and inpainting sparse and incomplete point clouds is a challenging task. Existing methods mainly focus on upsampling low-resolution point clouds or inpainting incomplete point clouds. In this paper, we propose a unified framework for upsampling and inpainting point clouds simultaneously. Specifically, we develop a point cloud upsampling and inpainting network called PUI-Net, which consists of the attention convolution unit and non-local feature expansion unit. In the attention convolution unit, the channel attention mechanism is adopted to extract discriminative features of point clouds. With the extracted discriminative features, we employ the non-local feature expansion unit to generate dense feature maps of high-resolution point clouds. Furthermore, we formulate a novel inpainting loss to train the PUI-Net so that the holes in the incomplete clouds can be completed. Quantitative and qualitative comparisons demonstrate that our proposed PUI-Net can yield good performance in terms of the upsampling and inpainting tasks on the incomplete point clouds.