Aerial3D: Real-word 3D Reconstruction for UAV Images

Image-based 3D reconstruction constitutes a prominent research topic in the field of computer vision, attracting considerable attention both domestically and internationally. The ubiquity of consumer-level unmanned aerial vehicles (UAV) has endowed us with a wealth of aerial images, fostering the emergence of an abundant source of image data for large-scale outdoor scene 3D reconstruction. This paper focuses on addressing the challenge of large-scale outdoor 3D reconstruction using aerial images. The proposed method, Aerial3D, employs an incremental structure-from-motion method for computing a sparse point cloud model and camera parameters. It further utilizes a multi-view stereo method based on depth map fusion for generating a dense point cloud model. The paper introduces a global view selection method, considering factors such as principal axis angle, image resolution, and timing difference between images, to enhance the accuracy and completeness of the point cloud model. The Floating Scale Surface Reconstruction (FSSR) method is adopted for creating mesh and textured models from the generated dense point clouds. The test results of Aerial3D in various real-world scenarios demonstrate that its performance surpasses that of traditional methods such as MVE and COLMAP.