Honeycomb-Surfels: Compact Hexagonal Primitives for Efficient Scene Rendering and Reconstruction

computer-vision3d-reconstructionnovel-view-synthesisgaussian-splattingrenderingcomputer-graphicsmachine-learningoutside

A 15-page preprint (v1 submitted 27 Oct 2026, 36.9 MB with figures) from Prakhar Kaushik and Soumava Paul (equal contribution) and Alan Yuille, all of the Computational Cognition, Vision, and Learning (CCVL) lab at Johns Hopkins University. It introduces "Hexels": planar hexagonal surface primitives with a linear ("first-order") per-primitive color field, pitched as a drop-in replacement for the elliptical Gaussian kernel used in Gaussian-splatting-family scene representations. The paper states a project page (honeycomb-surfels.github.io) and a code repository (github.com/honeycomb-surfels/hexels); neither was independently fetched this session. Both the arXiv abstract page and the PDF were fetched and archived cleanly (PDF fetch recorded tls:"verified"); no addressed-to-AI language, override language, tier self-assignment, credential requests, or other safety signals fired on either — no ## Safety flags section is needed for this capture.

Claim: Hexels are planar hexagonal surface primitives carrying a first-order (linear) per-primitive color field, built as a direct substitution for the Gaussian kernel inside the existing 2D Gaussian Splatting (2DGS) rasterization and depth-compositing pipeline

Claim type: specific technical-mechanism (how the system actually works, not merely that it exists). Source: the paper itself — Tier 1 (primary authors' own description of their own method), clearing the Tier 1–2 floor required for mechanism claims.

Each Hexel is centered at a 3D point with two orthogonal tangent directions and two scales, giving an oriented planar patch (mirroring how 2DGS parameterizes its surfels). Instead of a Gaussian falloff, the primitive's spatial support is a soft hexagonal boundary function — a signed-distance-like term fed through a sigmoid, with a sharpness parameter controlling how crisply the hexagon's edge is defined. Color is not constant across the primitive (as in standard Gaussian/triangle splats) but varies linearly from a base RGB value via two RGB gradient vectors, so a single primitive can represent smooth shading or color variation across its footprint. Critically, the paper describes the rendering/compositing machinery itself — perspective-correct ray–plane intersection, front-to-back depth-ordered alpha blending — as carried over unchanged from 2DGS; only the per-primitive weighting function and color model change.

source_quote: "We introduce Hexels: novel planar hexagonal surface elements with a first-order spatial color field for efficient 3D scene and 2D image representation." source_url: https://arxiv.org/pdf/2610.37668 source_sha: 386fba0b61a21c7a9a9a4549d93fb9d8fbb19d741878b8427dfb16c9554b36b3 source_tier: 1

source_quote_2: "Thus, relative to 2DGS, the rasterization and compositing remain unchanged; we replace the Gaussian weight with the Hexel support and evaluate spatially varying color within each primitive." source_url_2: https://arxiv.org/pdf/2610.37668 source_sha_2: 386fba0b61a21c7a9a9a4549d93fb9d8fbb19d741878b8427dfb16c9554b36b3 source_tier: 1

Claim: At matched model-memory budgets, Hexels improve PSNR over 2D Gaussian Splatting (2DGS) by 4.53–7.38 dB and over Triangle Splatting by 1.70–2.90 dB across three novel-view-synthesis benchmarks, while using less peak training VRAM

Claim type: quantitative (specific dB figures and benchmark results). Source: the paper itself — Tier 1, clearing the floor.

The comparison holds total uncompressed model memory fixed (so Hexels and the baselines are allowed different primitive counts) across the Tanks and Temples, Mip-NeRF360, and DTU benchmarks — 34 scenes in total. Under this constraint the paper reports Hexels (its "V1" variant) beating 2DGS by 4.53–7.38 dB of PSNR and Triangle Splatting by 1.70–2.90 dB, while requiring lower peak training VRAM and rendering up to 1.87× faster than 2DGS in the same setting.

source_quote: "At matched model-memory budgets, Hexels V1 improves PSNR over 2DGS by 4.53–7.38 dB and over Triangle Splatting by 1.70–2.90 dB across the three benchmarks." source_url: https://arxiv.org/pdf/2610.37668 source_sha: 386fba0b61a21c7a9a9a4549d93fb9d8fbb19d741878b8427dfb16c9554b36b3 source_tier: 1

Claim: At matched primitive-count budgets (same number of primitives as the baseline, rather than same memory), Hexels improve PSNR over 2DGS by up to 4.74 dB while using 1.76–1.95× less model size, up to 7× less peak training VRAM, and rendering up to 2.31× faster

Claim type: quantitative ("Nx" multipliers and specific dB figures). Source: the paper itself — Tier 1, clearing the floor.

This is a separate experimental condition from the matched-memory comparison above: here each scene's structure-from-motion point count sets a common primitive-count ceiling for every method, isolating the per-primitive design (hexagonal support plus linear color field) from any advantage that comes merely from using fewer, larger primitives at a fixed memory budget. Under this condition the paper still reports a meaningful PSNR gain over 2DGS, now paired with substantially lower per-primitive storage and training memory, and faster rendering.

source_quote: "At matched primitive budgets, Hexels improves PSNR over 2DGS by up to 4.74 dB, using 1.76–1.95× less model size and upto 7× less VRAM, and renders up to 2.31× faster (Table 3)." source_url: https://arxiv.org/pdf/2610.37668 source_sha: 386fba0b61a21c7a9a9a4549d93fb9d8fbb19d741878b8427dfb16c9554b36b3 source_tier: 1

Claim: For 3D surface reconstruction, Hexels cut mean Chamfer-distance error by roughly 28.9% relative to memory-matched 2DGS across 15 DTU scans (0.978 mm vs. 1.375 mm), and achieve the highest mean F1 score among compared methods on 6 Tanks and Temples surface-reconstruction scenes (34.38% vs. 31.72% for 2DGS)

Claim type: quantitative (specific error/percentage figures and a benchmark comparison). Source: the paper itself — Tier 1, clearing the floor.

Surface reconstruction is evaluated separately from the novel-view-synthesis comparisons above, using Chamfer distance (lower is better) on DTU scans and F1 score (higher is better) on Tanks and Temples surface-reconstruction scenes, both under the same memory-matched-vs-2DGS setup. The DTU figure (0.978 mm for Hexels' "V3" variant vs. 1.375 mm for 2DGS) implies the ~28.9% relative reduction the abstract states; the Tanks and Temples F1 comparison (34.38% vs. 31.72%) is reported directly against 2DGS's result in the paper's "unconstrained" settings table, with Hexels also beating Triangle Splatting (18.63%) and MeshSplatting (27.13%) in the same table.

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_url: https://arxiv.org/pdf/2610.37668 source_sha: 386fba0b61a21c7a9a9a4549d93fb9d8fbb19d741878b8427dfb16c9554b36b3 source_tier: 1

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

Further leads

• Hexels also outperform Image-GS [40] on 2D image (not 3D scene) representation — e.g. 33.16→35.82 dB PSNR, fewer training iterations (5.1K vs. 10K) — on a 45-image benchmark (Table 1); a distinct application claim from the 3D results above, not pursued further here. — source: arXiv:2610.37668 (PDF), §3.1 and Table 1.
• The choice of hexagon as the primitive's shape is explicitly grounded in Thomas C. Hales's honeycomb conjecture/theorem (regular hexagons tile a plane without gaps and need less perimeter than squares or triangles for equal area); the paper's own Appendix G derives that an optimally-oriented hexagon can cover "about 30% more area than the corresponding square and twice the area of the corresponding triangle for the same bounded color error." Worth a dedicated geometry/mechanism note once the appendix proof is read directly. — source: arXiv:2610.37668 (PDF), Introduction/§3 (citing ref [9], Hales, "The Honeycomb Conjecture," Discrete & Computational Geometry 25:1–22, 2001) and Appendix G (not independently read this session).
• The paper reports an ablation isolating shape from color: swapping a Gaussian footprint for a hexagonal one at fixed appearance model reduces "overdraw" (mean blended contributions per pixel) from 10.47 to 2.48 on a shaded test target while improving PSNR — a mechanism-level claim about why the hexagonal shape helps, distinct from the headline benchmark numbers above. — source: arXiv:2610.37668 (PDF), §4.4 and Figure 9.
• "Direct surface-cell extraction": the paper claims Hexels can be converted to an explicit polygonal mesh-like surface in O(1) work per primitive directly from the optimized representation, without the depth-fusion/Marching-Cubes pipeline 2DGS uses for mesh extraction — a downstream-applications claim not evaluated quantitatively against alternatives in the excerpt read. — source: arXiv:2610.37668 (PDF), §4.3.
• Future-work note: the authors propose "developable Hexels," curved (zero-Gaussian-curvature) variants of the primitive, as an unimplemented extension — a lead for tracking whether this gets built out in a later paper. — source: arXiv:2610.37668 (PDF), Limitations and Future Work.
• Project page (https://honeycomb-surfels.github.io/) and code repository (https://github.com/honeycomb-surfels/hexels), both named in the paper's own header, were not independently fetched or verified this session.

Entity candidates

• Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, Shenghua Gao — people/research group — authors of 2D Gaussian Splatting ("2D Gaussian Splatting for Geometrically Accurate Radiance Fields," ACM SIGGRAPH 2024 Conference Papers, 2024) — the FOUNDATIONAL prior method this entire paper measures itself against: nearly every headline number in this capture is a comparison against 2DGS, and the paper explicitly carries over 2DGS's rasterization/compositing pipeline unchanged.
• Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, George Drettakis — people/research group — authors of 3D Gaussian Splatting ("3D Gaussian Splatting for Real-Time Radiance Field Rendering," ACM Transactions on Graphics 42(4), 2023) — the deeper ancestor in the lineage (volumetric Gaussian primitives that 2DGS itself replaced with planar surfels, and that this paper also benchmarks against directly as "3DGS").
• Thomas C. Hales — person (mathematician) — author of "The Honeycomb Conjecture" (Discrete & Computational Geometry 25:1–22, 2001), the mathematical result the paper cites as justification for choosing a hexagon over a square or triangle as the primitive's shape.
• Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, Ren Ng — people/research group — authors of the original NeRF paper (ECCV 2020), the still-earlier foundational method (neural radiance fields) that the whole Gaussian-splatting/surfel lineage this paper sits in was built to replace for speed.
• Prakhar Kaushik — person — co-first author (Johns Hopkins CCVL lab).
• Soumava Paul — person — co-first author (Johns Hopkins CCVL lab).
• Alan Yuille — person — senior author (Johns Hopkins CCVL lab); a long-established computer-vision researcher whose name recurs across many CV subfields, potentially a vault entity worth tracking across papers rather than just this one.
• Hexels / Honeycomb-Surfels — concept/term — the paper's own named primitive; candidate for a standalone concept note once/if it recurs elsewhere in the vault.
• 2D Gaussian Splatting (2DGS) — concept/term — candidate for a standalone concept note given how load-bearing it is as this paper's baseline.
• 3D Gaussian Splatting (3DGS/Gaussian splatting) — concept/term — candidate for a standalone concept note as the root method of this entire lineage.
• Honeycomb conjecture — concept/term — the 2001 Hales theorem about optimal plane-tiling; a reusable mathematical concept independent of this specific paper.
• Image-GS — concept/term — the 2D-image-representation baseline (Zhang et al., SIGGRAPH 2025) this paper also compares against for its image-representation claims.