Faculty of Geomatics, East China University of Technology, Nanchang, Jiangxi 330013, China Abstract It is difficult to filter the data from crowd-sourced images, which leads to the geometric missing and noises in the point cloud data generated from the images. To solve this problem, a method of 3D reconstruction based on crowd-sourced images is presented. Firstly, the methods based on website Application Programming Interface (API) and web page analysis are used to obtain the crowd-sourced images. Then, the crowd-sourced images are filtered by deep learning to obtain high-quality crowd-sourced pictures data. Finally, the structure from motion (SFM) algorithm is used to complete the 3D reconstruction based on the filtered crowd-sourced images. It is concluded that image set screened by deep learning algorithm is more suitable for 3D reconstruction, so as to solve the disadvantages and deficiencies of crowd-sourced image, an emerging data source, in the application of 3D modeling.