Answers are grounded in this paper and cite their sources.
Over time, cultural heritage sites deteriorate, making it vital to document their original state for future generations. We present an efficient system for 3D modeling and reconstruction of cultural heritage, demonstrated on Patan Durbar Square. Short video clips were captured from various angles using an iPhone 13, and individual frames were extracted to form overlapping datasets of about 900–1300 frames per site. Using COLMAP for camera pose and sparse point cloud estimation, Gaussian splatting reconstructed both geometry and appearance. However, the baseline implementation introduced artifacts and required extensive training. We therefore evaluated multiple strategies, including Splatfacto and artifact removal and segmentation methods. Combining GroundingDINO and SAM effectively removed people and birds, while Splatfacto with masking and bilateral grids produced high-quality reconstructions. The proposed pipeline enables accurate, scalable, and efficient 3D documentation of cultural heritage sites.
No takes yet. Share an insight, caveat, or question.
Shrestha et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: