Towards efficient 3D Gaussian Splatting reconstruction with PairVPR-based image selection: A glacier UAV case study

Large-scale 3D reconstruction from UAV imagery is essential for remote sensing and environmental monitoring, yet modern neural rendering methods such as 3D Gaussian Splatting (3DGS) are computationally intensive due to redundant views. Reducing input images while preserving reconstruction accuracy and completeness remains a practical challenge. We present a scalable image subset selection framework that integrates transformer-based visual place recognition (PairVPR) with a facility-location selection strategy to identify and remove redundant views while maintaining spatial coverage. We first validate the approach on six standard benchmark scenes (five Mip-NeRF360 scenes and the Tanks&Temples Truck scene): under a fixed image budget, PairVPR-based selection attains the best average PSNR, SSIM, and LPIPS among all evaluated VPR selectors. Building on this, we present a case study on a real-world glacier UAV survey of 589 images, in which pruning to a compact subset reduces overall runtime by more than half at the most aggressive setting, while incurring only a modest loss in reconstruction quality, and in which PairVPR-based selection again attains the best reconstruction quality among the evaluated VPR selectors. By combining learned similarity with a coverage-driven selection process, the method operates with linear memory complexity and adapts to large image collections, offering a practical solution for large-scale geospatial mapping and monitoring workflows.