A Memory-Efficient SfM Approach for Large-Scale UAV Imagery Scene

With the rapid development of the low-altitude economy, UAV photogrammetry has emerged as a critical spatial data acquisition technology in fields such as energy infrastructure inspection, emergency response, and topographic mapping. However, the inherent characteristics of UAV imagery—low flight altitude and high overlap—pose significant computational efficiency challenges for Structure from Motion (SfM) processing in large-scale scenes. To address this, we propose a memory-efficient SfM method for large-scale UAV image datasets. Our approach consists of three key phases: initially, a divide-and-conquer strategy partitions the UAV imagery into block units based on feature matching relationships, ensuring sufficient overlap between adjacent blocks. Subsequently, localized SfM reconstruction is implemented within each individual block. Finally, leveraging pose consistency constraints between overlapping blocks enables the effective merging of multiple SfM blocks. Experimental results demonstrate that the proposed method achieves large-scale UAV imagey SfM processing under constrained memory conditions while maintaining computational viability.