For 3D surface reconstruction, Hexels cut mean Chamfer-distance error by ~28.9% versus memory-matched 2D Gaussian Splatting on DTU, and post the highest mean F1 score among compared methods — including unconstrained 2DGS — on Tanks and Temples

computer-visiongaussian-splatting3d-reconstructionbenchmarksoutside

Surface reconstruction is evaluated separately from the novel-view-synthesis comparisons in the matched-memory PSNR note, using Chamfer distance (lower is better) on 15 DTU scans and F1 score (higher is better) on 6 Tanks and Temples surface-reconstruction scenes. The Chamfer comparison is under the same memory-matched-vs-2D Gaussian Splatting setup used elsewhere in the paper: on DTU, the paper's "V3" Hexel variant reports a mean Chamfer distance of 0.978 mm against 1.375 mm for memory-matched 2DGS — a reduction the paper's own abstract states as roughly 28.9% relative (the paper's Table 5 also shows unconstrained 2DGS ahead of Hexels there, at 0.794 mm, with roughly 3× Hexels' peak training VRAM). The F1 comparison is scoped differently: the 31.72% that Hexels' 34.38% is quoted against is 2DGS's unconstrained figure in Table 6 — 2DGS at its best setting, not memory-matched; memory-matched and primitive-matched 2DGS score 22.05% and 21.03% in the same table, so the gap against the budgeted variants is wider. Hexels also beat Triangle Splatting (18.63%) and MeshSplatting (27.13%) in Table 6 — though this note's claim is scoped to the 2DGS comparisons that both source quotes directly ground.

That geometry-quality gains (Chamfer, F1) and rendering-quality gains (PSNR, in the companion notes above) move together is notable: it suggests the hexagonal support and linear color field aren't merely fitting training views more precisely at the expense of underlying geometry, but are reconstructing a more accurate surface too. The paper offers no independent corroboration of either figure yet.