Evaluating Visual Performance of Multi-object Scene Reconstruction using Triangulation from Airborne Geo-spatial Point Clouds

This paper presents a hybrid framework for reconstructing the scene of objects from a geospatial dataset from point clouds. The method addresses the processing issues of large data points in 3D space and uses spatial clustering to determine the context membership for points. It further uses the min-cut based graph algorithm to segment the resulting polyhedral surface of the object in the scene. Spatial clustering is performed using k-dimension search tree in multidimensional space for segmenting and assigning memberships of point to possible clusters. Size of these clusters controls the background penalties for use during graph cuts for generating multi-object polyhedral objects. The results in this paper are compared to other methods and it is observed that the min-cut graph-based approach outperforms simple Delaunay triangulation and Euclidean estimate based method for surface construction. It can classify objects with better accuracy into separated structures and enabling more efficient visual communication and rendering performance. Geospatial reconstruction problems especially those used for visual communication can use the min-cut based method to construct multi-object scenes for faster rendering performances and reduced data management overheads.