ScholarWorks@성균관대학교

초록

Large-scale outdoor scenes captured by drones exhibit wide spatial extent and diverse viewpoint distributions, making training view density an important factor in view synthesis. While recent Gaussian Splatting-based methods have demonstrated efficient real-time rendering, their behavior under varying numbers of training views in real-world, building-scale scenes remains insufficiently explored. This work introduces a large-scale drone-captured dataset of a university library and its surrounding area, which has been adopted as a representative large-scale outdoor sequence in the MPEG GSC common test conditions (CTC) to support standardized evaluation of Gaussian Splatting-based reconstruction in outdoor environments. The dataset was captured along circular flight trajectories at multiple altitudes and distances, providing systematic coverage of near-range and far-range viewpoints. Using a fixed set of evaluation views, the results show that faithful large-scale scene coverage can be achieved with as few as approximately 200 training images, despite the wide spatial extent of the outdoor environment.

키워드