Geometry-complete benchmark for completion evaluation

Develop a geometry-complete benchmark containing realistic synthetic or equivalent scenes with complete ground-truth surfaces for quantitatively evaluating completion-based 3D reconstruction methods.

Background

The paper identifies the lack of suitable evaluation data as a direct obstacle to measuring surface completion. Existing sensor-reconstructed meshes, including those in ScanNet++, omit or incompletely represent surfaces hidden by objects, so plausible completed geometry is incorrectly counted as false-positive geometry.

The authors state that a geometry-complete ground truth is necessary and that such data are unavailable at the required scale and complexity. Creating this benchmark would enable fair quantitative comparisons of background and amodal completion methods.

References

Fully validating the fidelity of completed surfaces requires a geometry-complete ground truth---such as realistic, synthetic CAD-modeled environments---which does not currently exist at the scale and complexity of ScanNet++. We view the development of such complete synthetic benchmarks as an essential prerequisite for the quantitative evaluation of completion-based reconstruction techniques.

— ObjectSplat: Improving Mesh Fidelity and Interactivity for 3D Scenes via Object-Level Mesh Splatting  (2608.30423 - Kamal et al., 31 Aug 2026) in Section 4, “Limitations and Future Work”; Appendix, Section “Surface Completion”