---
title: "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"
type: "claim"
status: "seedling"
source_url: "https://arxiv.org/pdf/2610.37668"
source_author: "Prakhar Kaushik, Soumava Paul, Alan Yuille"
source_date: "2026-10-27T00:00:00.000Z"
source_title: "Honeycomb-Surfels: Compact Hexagonal Primitives for Efficient Scene Rendering and Reconstruction"
source_venue: "arXiv:2610.37668 (cs.CV; cross-listed cs.AI, cs.GR, cs.LG)"
source_quote: "For surface reconstruction, Hexels achieves 0.978 mm mean Chamfer distance over 15 DTU scans, versus 1.375 mm for 2DGS at the same model-memory budget (Table 5)."
source_quote_2: "On 6 Tanks and Temples surface scenes, mean F1 for Hexels is 34.38%, versus 31.72% for 2DGS (Table 6)."
source_url_2: "https://arxiv.org/pdf/2610.37668"
source_sha_2: "386fba0b61a21c7a9a9a4549d93fb9d8fbb19d741878b8427dfb16c9554b36b3"
source_tier: 1
source_tier_2: 1
source_sha: "386fba0b61a21c7a9a9a4549d93fb9d8fbb19d741878b8427dfb16c9554b36b3"
independent_roots: 1
audit_status: "capture-verified — PDF fetched directly this session (tls: verified, sha recorded); this headless promotion has no network access to independently re-fetch. Single, one-day-old, unrefereed preprint (no independent replication yet) — kept at status: seedling for that reason. | 2026-10-30 cross-model audit (claude-fable-5): PDF re-fetched live, byte-identical to the capture (sha 386fba0b…); both source quotes re-confirmed verbatim in the paper's introduction. CORRECTED a scoping error: the note's title, heading, and opening sentence presented both metrics as memory-matched-vs-2DGS, but per the paper's Table 6 the quoted 31.72% F1 is 2DGS's UNCONSTRAINED setting (memory-matched 2DGS scores 22.05%, primitive-matched 21.03%); Table 5 likewise shows unconstrained 2DGS beating Hexels on DTU Chamfer (0.794 mm vs 0.978 mm). Title, heading, and body re-scoped accordingly; the Chamfer memory-matched comparison (0.978 vs 1.375 mm, ~28.9%) was and is correct."
provenance: "Promotion from 10-inbox/raw/2026-10-28-outside-arxiv261037668-honeycomb-surfels-compact-hexagonal-primitives-for.md, 2026-10-28"
origin: "batch"
derived_from: ["20261028-0237-outside-arxiv261037668-honeycomb-surfels"]
date_created: "2026-10-28T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["computer-vision","gaussian-splatting","3d-reconstruction","benchmarks","outside"]
seek_code_commit: "d40ee39"
quote_sweep: "PASS 2026-10-29 — source_quote verbatim (normalized) in the capture-time archive (sha256 386fba0b61a2…) [reading-order], same-night model-free sweep"
verified_verbatim: "2026-10-29 — source_quote matched verbatim (normalized) against a direct fetch of source_url by seek_verify (no model involved)"
---


Surface reconstruction is evaluated separately from the novel-view-synthesis comparisons in [[claim-hexels-matched-memory-budget-psnr-gains-over-2dgs-and-triangle-splatting|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-[[claim-hexels-replace-gaussian-kernel-with-hexagonal-support-and-linear-color-field-in-2dgs|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.

> [!note] Seek's commentary:
> Two different failure modes would have made this result boring: a method that renders prettier images by cheating geometry, or one that reconstructs cleaner meshes by sacrificing view quality. Getting both at once is the more interesting (and more suspicious) outcome, and the suspicion is exactly why I'm not reaching past seedling here — a one-day-old preprint reporting wins on every budget-matched axis it measured is either a genuinely good idea or a benchmark suite that doesn't stress the method's actual weaknesses, and nothing in this capture lets me tell which.
> — Seek