MANAT: A Filtering-Based Method for Denoising Nonuniform Photogrammetric Point Clouds

Abstract

Three-dimensional point clouds reconstructed from photogrammetry often exhibit noise and non-uniform sampling density, which challenges existing denoising methods that rely on precise normal estimation or extensive parameter tuning. This study presents the Multi Attribute Neighbour Attraction Technique (MANAT), a novel single-stage, density-adaptive filtering method that jointly leverages spatial position, surface normals, and color as inherent photogrammetric attributes for unified noise removal. MANAT assesses each point’s consistency within its k-nearest neighbourhood using local geometric, orientation, and color statistics, enabling effective discrimination between valid surface points and noise in real-world photogrammetric data. On a large-scale heritage dataset of 141.7 million points, MANAT achieved 23.78% noise removal with improvements of 9.60, 6.91, and 4.40% in surface roughness, local and global normal standard deviations respectively. Comparison with DBSCAN confirms that spatial density alone is insufficient to characterise embedded photogrammetric noise, highlighting the necessity of multi-attribute denoising. These results demonstrate MANAT’s practical effectiveness as a robust framework for enhancing the accuracy and reliability of photogrammetric 3D reconstructions under realistic acquisition conditions.