Abstract: The study of 3D reconstruction of single-view image point clouds with deep learning has recently become one of the hot topics in the field of computer vision. In view of the remarkable results achieved by some methods in this direction, a more comprehensive summary of the work on the subject in recent years is intended to provide a reference for researchers in this field, and at the same time serve as an introduction for researchers who are interested in this method and further advance the existing research status. The literature was organized according to the classification of different point cloud representations, and then the basic ideas, training mechanisms, learning paradigms and relationships between algorithms of the corresponding methods were reviewed. The common datasets, loss functions and evaluation methods in the field were discussed, and their characteristics, limitations and download addresses were summarized and sorted out. In addition, the performance of some important methods on public datasets was analyzed and compared from the perspective of network structure and supervision. We summarized some remaining problems and discussed possible future development trends by combing the current research status.