Three-dimensional (3D) reconstruction technology can be applied in many fields such as 3D CT images in medicine,path planning in robots,and display of relics in archaeological work.The completeness and accuracy of 3D reconstruction are directly related to steps of Structure from Motion (SfM) and Multi-View System (MVS).In this paper,we introduce these two steps in detail.In the reconstruction of sparse point cloud,feature point detection and matching and SfM reconstruction methods are introduced and elaborated.We summarize the researches on feature point detection and matching in recent years,and classify the reconstruction methods of SFM into global and incremental ones,and introduce them in detail.In the dense point cloud reconstruction,the depth map estimation is introduced and elaborated,in which the traditional geometric calculation method,hybrid geometry-and deep learning-based approach,and deep learning based method are summarized.Meanwhile,the evaluation index of 3D reconstruction is introduced.Finally,the image based 3D reconstruction is summarized,and the problems facing the 3D reconstruction and possible future research directions are listed.