Abstract: Structure from motion (SfM) aims to compute the absolute camera poses in a global unified coordinate system based on local feature matches between image pairs, which is a key problem in image-based 3D reconstruction. In recent years, with the development of acquisition devices, computational resources, and theoretical methods, research on SfM has gradually expanded from small-scale controlled laboratory settings to large-scale real-world indoor and outdoor environments, by which significant progress in both theoretical methods and practical applications has been achieved. Starting from the perspective of practical application, this survey focuses on the latest advances in large-scale 3D scene reconstruction-oriented SfM research, and compared with existing surveys in the related field, it specifically concentrates on the core problem of camera pose estimation in SfM and provides a comprehensive overview of the latest developments on both analytical and learning-based approaches. On this basis, to facilitate community development, the current progress and future trend in SfM research are also discussed and analyzed by this survey.