It is nice to see a method that can reconstruct a 3D scene using sparse input instead of dense input to achieve better performance for computation time and spatial complexities.
In this paper, the authors present a method for reconstructing a manifold surface from the sparse structure-from-motion data. Based on Harris points, a 3D Delaunay triangulation is calculated, which encodes the potential adjacencies between the points. Ray tracing is then used to label the tetrahedra. A 2-manifold extraction is computed by region growing, using a greedy algorithm. Post-processing--including peak removal, surface denoising, sky removal, and spurious handle removal--helps achieve the final result. The time complexity is studied with a table that clearly shows the worst-case time complexities for each step. The experiments are conducted on both still and video image sequences, and the system demonstrates good results.
The authors point out that “the current system is able to reconstruct the main components of environments (ground, building, dense vegetation, etc.), but the main limitation is due to the lack of points to reconstruct thin scene details.”
This paper is an extended version of the authors’ previous work [1,2]. I strongly recommend that those interested in this approach read the other papers [1,2] before jumping into this one.