Evaluating Surface Mesh Reconstruction Using Real Data

Surface reconstruction has been studied thoroughly, but very little work has been done to address its evaluation. In this article, we propose new visibility-based metrics to assess the completeness and accuracy of three-dimensional meshes based on a point cloud of higher accuracy than the one from which the reconstruction has been computed. We use the position from which each high-quality point has been acquired to compute the corresponding ray of free space. Based on the intersections between each ray and the reconstructed surface, our metrics allow evaluating both the global coherency of the reconstruction and the accuracy at close range. We validate this evaluation protocol by surveying several open-source algorithms as well as a piece of licensed software on three data sets. The results confirm the relevance of assessi ng local and global accuracy separately since algorithms sometimes fail at guaranteeing both simultaneously. In addition, algorithms making use of sensor positions perform better than the ones relying only on points and normals, indicating a potentially significant added value of this piece of information. Our implementation is available at https://github.com/umrlastig/SurfaceReconEval.