Point-Reconstruction-Network: 3D Point Cloud Reconstruction from a Single Image

Reconstructing the 3D shape of the object from the single-view image has far-reaching applications in medical research, digital protection of cultural relics, and game software development. With the recent proliferation of deep learning, various deep models have been proposed for the 3D point cloud generation. However, these networks generally have problems such as serious loss of features during network propagation, poor reconstruction quality and large amount of network parameters. In order to solve this problem, this paper proposes a point cloud-based neural network named point-reconstruction-network. The network uses the fine-tuned residual network to extract the features of the input RGB image and gets the feature vector Z. Then directly converts it into a point cloud output through the generator, and obtains the 3D point cloud shape of the object. The point-reconstruction-network can reconstruct the 3D point cloud shape for the RGB image with sufficient information. The proposed network is evaluated on ShapeNet and Pix3D datasets, and the results show that proposed approach outperform the classic approaches on the task of single-view 3D object reconstruction.