{"id":"b597275e-bd64-4fbf-9abb-6aa84a05df81","arxiv_id":"2505.20267","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A dual representation using learnable triangles plus neural Gaussians achieves competitive rendering and more compact geometric abstractions on common 3D scene benchmarks.","lead":"HaloGS is a 3D scene representation that pairs lightweight triangles for geometry with neural Gaussians for appearance, and it extracts compact planar meshes from the triangles. It matters because it offers a way to keep photorealistic rendering while also producing small, editable geometry for indoor and outdoor scenes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Geometry-accuracy claim rests on a planar-extraction comparison that may favor triangle-soup inputs; per-method parameter tuning is needed before the Table 3 gap is credible.","rationale":"The reader's concern about monocular depth and normal priors is reasonable, and the fine-stage self-supervision in Eq. (6) does create a closed loop that could lock in prior errors. However, the paper's own ablation of the fine-stage Gaussian-based refinement shows only a small PSNR drop while preserving most of the triangle geometry, suggesting the priors are not the only support for the geometric result. The more decisive vulnerability is the quantitative geometry evaluation: Table 3 is the primary evidence for the 'accurate geometry' half of the central claim, and its methodology does not control for the radically different input formats fed to the same planar extractor. If the comparison is biased, the central claim loses its quantitative backing even if the training pipeline works exactly as described. A per-method parameter search and a controlled input-format experiment would settle this directly. The rendering comparisons in Tables 1-2 show HaloGS is often on par with, rather than clearly better than, the strongest baselines, so the geometry evaluation is the differentiator and must be trustworthy. Because the concern is addressable with additional experiments rather than a demonstrated flaw, the reader's CONDITIONAL verdict remains appropriate.","tokens_in":20347,"tokens_out":6216,"duration_ms":66905,"concrete_test":"Re-run the geometry evaluation of Table 3 with a per-method grid search over GoCoPP parameters (ϵ, σ, thn), reporting the Pareto front of Chamfer distance versus number of planar primitives for each method. Additionally, feed the same triangle-to-point sampling routine used for Ours into 2DGS's extracted mesh before planar extraction, thereby controlling for input format. If, under per-method tuning, any baseline achieves Chamfer distance within noise of Ours at comparable primitive count, the claimed geometric advantage does not hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline geometry claim (Sec. 4.2, Table 3) is established by extracting planar shapes with GoCoPP from Ours (oriented points sampled from perfectly planar triangles) and from baseline outputs (MVS point clouds, 2DGS meshes, GFSGS point clouds), using the same fixed 10-iteration coarse-to-fine parameter schedule (Table 5). GoCoPP is sensitive to input density, noise, and normal coherence, so the comparison conflates representation quality with input-format quality. The triangle-soup input is nearly ideal for planar clustering, while baseline point clouds and meshes contain noise and irregular sampling that the fixed schedule may over- or under-segment. The per-scene primitive counts for Ours in Table 11 vary widely (e.g., 1039 to 2902 on ScanNet++), suggesting extraction settings dominate the reported numbers. Separately, Eq. (6) contains a likely typo (W⊙Dgs is used twice where Dgs and Ngs are intended), and the Sec. 5 promise of a training-schedule sensitivity analysis in the Supplementary is not fulfilled; these are secondary but weaken reproducibility. If the Table 3 gap reflects extraction parameters rather than genuine geometric superiority, the central claim that HaloGS provides 'compact, accurate geometry' is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HaloGS, a dual representation that decouples geometry from appearance: coarse learnable triangle primitives provide explicit, compact geometry, while neural Gaussians decoded from triangle features are responsible for photorealistic rendering, organized in an Octree-GS-style level-of-detail structure. Training is coarse-to-fine: the coarse stage supervises triangles with monocular depth and normal priors (Depth-Anything-V2 and StableNormal, after RANSAC alignment with SfM points), and the fine stage adds photometric supervision together with Gaussian-rendered depth and normal maps as pseudo-ground truth for further geometric refinement. Adaptive density control (edge splitting, triangle pruning, opacity entropy regularization) is used to obtain clean reconstructions. Planar primitives are extracted from the optimized triangle soup with GoCoPP and stitched into compact meshes. Experiments cover 32 scenes from Deep Blending, ScanNet++, VR-NeRF, Zip-NeRF, MatrixCity, and FAST-LIVO2, with rendering metrics and geometric evaluations against ground-truth meshes and point clouds.","tokens_in":20659,"tokens_out":5924,"duration_ms":54475,"significance":"If the central claims hold, HaloGS is a useful point in the design space of hybrid scene representations: it offers an explicit low-poly geometry scaffold with photorealistic Gaussian appearance, compact storage, and downstream mesh usability, and its custom differentiable triangle rasterizer with subdivision and visibility correction is a concrete engineering contribution. Strengths include the broad evaluation across diverse indoor and outdoor datasets, per-scene result tables, and clean ablation studies for the adaptive density control and geometric detail enhancement. However, the headline geometry comparison is confounded by the planar-extraction protocol, the rendering results are not consistently above baselines, and a promised sensitivity analysis is missing from the supplementary material. The present evidence supports HaloGS as competitive and promising rather than as a demonstrated state-of-the-art improvement.","major_comments":[{"comment":"The geometry comparison is confounded by the planar-extraction protocol. GoCoPP is applied to all outputs with the same fixed 10-iteration coarse-to-fine schedule (Table 5), but the method is sensitive to input density, noise, and normal coherence. Uniformly sampled points from perfectly planar triangles are a near-ideal input, whereas MVS point clouds and 2DGS or GFSGS outputs have different noise and sampling characteristics, so the reported Chamfer distances conflate representation quality with input-format quality. The per-scene primitive counts in Table 11 (e.g., Ours ranges from 1039 to 2902 on ScanNet++) further suggest that extraction settings dominate the reported numbers. Please add a direct geometric evaluation (e.g., mesh or point-cloud Chamfer distance against ground truth without planar abstraction) and/or per-method tuning of extraction parameters, and report how Table 3 changes when the GoCoPP schedule is varied.","section":"§4.2, Table 3"},{"comment":"The claim that HaloGS 'consistently outperforms baselines in rendering quality' is not supported by the reported aggregates. In Table 2, FAST-LIVO2 PSNR is 29.81 for Ours versus 29.96 for 3DGS; in Table 1, VR-NeRF SSIM is 0.916 for Ours versus 0.918 for 3DGS. The per-scene tables (e.g., Tables 8–9) also show many scenes where Ours is below 3DGS or Octree-GS on one or more metrics. No error bars, repeated-run statistics, or significance tests are provided. Please temper the wording and add variance or significance analysis to support any comparative claims.","section":"§4.1, Tables 1–2; §A.2.1"},{"comment":"Equation (6) appears to contain a typo: the first Lgeo term writes W⊙Dgs twice, where the second argument should almost certainly be the Gaussian-rendered normal map W⊙Ngs. In addition, the wavelet-based weight map W is not fully specified (wavelet type, level, normalization), and its complement notation W=1−W is confusing. Because this loss is central to the fine-stage geometry refinement, please correct the equation and specify the weight map construction precisely.","section":"§3.3, Eq. (6)"},{"comment":"Section 5 promises that 'sensitivity to the chosen coarse-to-fine training schedule' is analyzed in the Supplementary, but Supplementary A.3 contains only qualitative limitations about transparent objects and sky/background regions, with no schedule-sensitivity analysis. This is a missing promised support. Given that the coarse stage relies on monocular priors (Eq. 5) and the fine stage uses Gaussian-rendered depth and normal maps as pseudo-labels (Eq. 6), please either provide the promised sensitivity analysis or remove the promise, and add an ablation that substitutes a different depth/normal prior (or ablates the RANSAC alignment) to show the geometric results are not tied to the specific priors.","section":"§5 and §A.3"}],"minor_comments":[{"comment":"Typo: 'Benifiting' should be 'Benefiting'.","section":"§1"},{"comment":"Typo: 'Aseembling' should be 'Assembling'; the same typo appears elsewhere in the figure.","section":"Fig. 1 caption"},{"comment":"Typo: 'traingles' should be 'triangles'; also 'shapeness and planeness' is awkward and should be rephrased.","section":"§3.2"},{"comment":"Zip-NeRF is described as an 'indoor-outdoor dataset' but its scenes are unbounded outdoor scenes; please correct the dataset description.","section":"§2, Table 2"},{"comment":"The ablation row 'w/o enhancing geomtirc detail using Gaussians' contains a typo ('geomtirc' should be 'geometric').","section":"Table 4"},{"comment":"Please add a reproducibility statement indicating whether training or evaluation code will be released.","section":"§4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper builds heavily on the authors' own Scaffold-GS, Octree-GS, and GoCoPP; the novelty lies in the integration, the triangle rasterizer, and the adaptive density control, but the evaluation would be more convincing with independent baseline implementations and a third-party validation of the geometry claims. The geometry comparison in Table 3 is the load-bearing evidence for the 'compact, accurate geometry' claim, and it is currently confounded by the planar-extraction input format. If the authors correct the Eq. (6) typo, deliver the promised sensitivity analysis, and re-run or supplement the geometry evaluation, the paper could become suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a credible engineering contribution. The new arrangement is a learnable triangle soup for geometry, loosely coupled with Scaffold-GS-style neural Gaussians for appearance, with adaptive triangle splitting/pruning and a LoD planar-extraction pipeline for compact meshes. The triangle splatting kernel with edge-preserving weights is a real piece of work. The two-stage training that first supervises triangles with monocular depth/normals and then refines them with Gaussian-rendered depth/normal feedback is sensible, and the ablations support the individual design choices.\n\nThe soft spots are mostly about the geometry evaluation. Table 3 extracts planar primitives from Ours and baselines using the same GoCoPP parameters, but the inputs are not the same kind of object: Ours feeds oriented points sampled from perfectly flat, surface-aligned triangles; MVS and GFSGS feed noisy point clouds; 2DGS feeds a fused mesh. GoCoPP is sensitive to density, noise, and normal coherence, so the fixed parameter schedule can benefit the cleanest input regardless of whether the underlying representation is geometrically better. The per-scene primitive counts for Ours vary widely, which hints the extraction settings dominate. I'd want to see per-method tuning or a sensitivity analysis before believing the factor-of-two Chamfer gap on ScanNet++.\n\nOther issues are minor: no error bars anywhere, and on a few metrics Ours is behind 3DGS (FAST-LIVO2 PSNR 29.81 vs 29.96; VR-NeRF SSIM 0.916 vs 0.918). Eq. (6) has a typo—W⊙Dgs appears twice where the second should be W⊙Ngs. And the conclusion promises a detailed sensitivity analysis in the Supplementary that I didn't find; the supplementary has implementation details and per-scene tables but no such analysis. None of these are load-bearing, but they weaken reproducibility.\n\nThe self-citation pattern is worth a note: the pipeline leans on Scaffold-GS, Octree-GS, and GoCoPP, all from the same group. That's not circular—those components are independently benchmarked—but it does mean the novelty is a combination rather than a fresh start.\n\nWho this is for: people building hybrid geometry+appearance representations for AR/VR or robotics, where a compact mesh coexists with photoreal rendering. They'll want to read the method section carefully and probably rerun the geometry comparison before citing the numbers.\n\nRecommendation: send it to peer review. The central idea deserves referee time, and the concerns I listed are addressable with an honest re-evaluation of the geometry comparison, error bars, and a code release. I would not accept the current geometry claim as-is.","headline":"A credible hybrid geometry-appearance system, but the headline geometry win is partly an artifact of comparing planar extraction on differently formatted inputs.","tokens_in":21205,"tokens_out":2795,"would_cite":true,"duration_ms":26151,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"HaloGS pairs learnable triangles with neural Gaussians to get compact meshes and photorealistic views in one model.","keywords":["3D Gaussian Splatting","triangle splatting","dual representation","compact mesh extraction","planar primitives","monocular depth and normal priors","novel view synthesis","level-of-detail rendering"],"falsifier":"Reconstruct a textureless white room or a low-light corridor with known ground-truth geometry, where monocular depth and normal priors are unreliable and SfM points are sparse. If HaloGS's triangle soup and compact mesh show Chamfer distance no better than 2DGS or MVS while its rendering PSNR stays high, the claim that accurate geometry and photorealistic appearance are jointly achieved would fail: appearance could be right while geometry is wrong.","tokens_in":20152,"feed_emoji":"📐","tokens_out":7445,"duration_ms":72266,"temperature":0.7,"pith_summary":"This paper tries to show that a 3D scene can be represented as a set of learnable triangles that carry the geometry plus a cloud of neural Gaussians that carry the appearance, and that this loose split avoids the usual trade-off between compact, editable geometry and photorealistic rendering. The authors evaluate the method on 32 indoor and outdoor scenes and report lower Chamfer distance to ground-truth geometry than MVS and 2DGS baselines while keeping rendering quality on par with or above 3DGS-based methods. If the claim holds, one trained model provides a low-poly mesh for storage, editing, and collision detection, and Gaussian splats for real-time novel views. The key is that geometry and appearance are optimized in separate, mutually supportive stages rather than entangled in a single representation.","feed_headline":"Keep geometry low-poly, rendering photoreal with one dual model","feed_subtitle":"Learnable triangles carry shape; neural Gaussians carry color, so one model gives editable meshes and real-time views.","key_machinery":"The load-bearing object is the learnable triangle primitive: three movable vertices parameterized through a local tangent frame (Eq. 2), rasterized by a custom CUDA kernel that computes ray–triangle intersections and blends contributions with an edge-preserving kernel (Eq. 3) to render depth and normal maps. A coarse stage supervises these triangles with monocular depth from Depth-Anything-V2 and normals from StableNormal, aligned to SfM points by RANSAC; a fine stage decodes neural Gaussians from each triangle following the Scaffold-GS scheme, embeds them in an Octree-GS octree, and supervises the triangles through Gaussian-rendered depth/normal maps weighted by a wavelet edge map, alongside image losses and adaptive splitting and pruning of triangles. The optimized triangle soup is clustered by GoCoPP planar primitive detection into multi-scale LoD planes that seed both rendering and the assembly of a compact mesh.","core_discovery":"HaloGS's central claim is that decoupling geometry from appearance, instead of fusing them into one implicit field or forcing Gaussians onto a thin surface, makes a representation simultaneously compact and photorealistic. The paper argues that learnable triangle primitives, splatted through a custom rasterizer with an edge-preserving kernel, can be driven to accurate surfaces by monocular depth and normal priors in a coarse stage, and that neural Gaussians decoded from those triangles then render images while feeding finer depth and normal maps back into the triangles in a fine stage. This loose coupling, the authors claim, is why the extracted planar primitives and compact mesh beat dense reconstructions in both fidelity and simplicity on planar indoor and urban scenes.","pith_inferences":["If the monocular priors are systematically wrong, such as on mirror-like or highly reflective surfaces, the coarse stage may lock triangles onto a plausible-but-wrong surface; a worthwhile test is comparing mesh quality against ground-truth LiDAR in regions where the depth prior and SfM disagree.","The same loose-coupling idea might transfer to other appearance decoders, such as a lightweight radiance field or feature grid, suggesting the triangle scaffold, not the specific Gaussian decoder, is the core contribution.","The wavelet-weighted fine-stage loss predicts that high-frequency texture regions receive more geometric refinement; an ablation that removes the weight map would reveal whether it chiefly improves edges or also affects flat-surface accuracy."],"forward_implications":["A trained HaloGS model gives both a photorealistic renderer and an editable low-poly mesh from the same data, so storage and real-time streaming can use the mesh while rendering uses the Gaussians.","In textureless or low-light areas where SfM points are sparse, the monocular-prior-guided triangles provide a coherent surface that keeps rendering quality from collapsing.","Because the triangles are explicitly optimized for geometry and only loosely coupled to appearance, the planar abstraction step can discard redundant geometry without hurting novel-view quality.","The compact mesh and planar segments can serve downstream tasks such as collision detection, geometric editing, and semantic reasoning directly from the trained model."],"supporting_citations":[{"why":"Establishes 3D Gaussian Splatting as the appearance baseline and the origin of the Gaussian primitives that HaloGS decodes from triangles.","marker":"[12]"},{"why":"Supplies the anchor-based neural Gaussian decoding scheme and the L_scaling regularization that HaloGS adapts for its fine stage.","marker":"[20]"},{"why":"Provides the ray–splat intersection and depth/normal rendering formulation that HaloGS's triangle rasterizer builds on, plus a geometric baseline.","marker":"[9]"},{"why":"Supplies the monocular depth prior (Depth-Anything-V2) used in the coarse-stage geometry loss.","marker":"[38]"},{"why":"Supplies the monocular normal prior (StableNormal) used in the coarse-stage geometry loss.","marker":"[39]"},{"why":"Contributes the octree LoD structure that embeds the neural Gaussians for level-of-detail rendering.","marker":"[25]"},{"why":"GoCoPP planar primitive extraction converts the triangle soup into the LoD planes and compact mesh.","marker":"[41]"},{"why":"The explicit ray–splat intersection algorithm is the geometric kernel of the custom triangle rasterizer.","marker":"[30]"},{"why":"Supplies the RANSAC-based alignment that rescales monocular depth to metric SfM coordinates.","marker":"[13]"},{"why":"Contributes the edge-preserving weight kernel (Eq. 3) and serves as a comparison baseline.","marker":"[8]"}],"fun_headline_variants":["Loose coupling of triangles and Gaussians for compact 3D scenes","HaloGS: Coarse triangles for geometry, Gaussians for appearance","One model, two primitives: low-poly shape, high-fidelity looks","Coarse meshes plus Gaussian splats for compact, adaptive 3D","Triangles carry shape, Gaussians carry color: HaloGS approach"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline stands on the assumption that monocular depth and normal predictions from pretrained diffusion models, after RANSAC alignment to sparse SfM points, are accurate enough to steer the triangle soup in the coarse stage, and that the fine stage's self-supervision from Gaussian-rendered depth and normals does not bake those priors' mistakes into the final geometry.","fun_headline_variants_meta":{"raw":{"variants":["Loose coupling of triangles and Gaussians for compact 3D scenes","HaloGS: Coarse triangles for geometry, Gaussians for appearance","One model, two primitives: low-poly shape, high-fidelity looks","Coarse meshes plus Gaussian splats for compact, adaptive 3D","Triangles carry shape, Gaussians carry color: HaloGS approach"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000198,"raw_usage":{"total_tokens":1308,"prompt_tokens":824,"completion_tokens":484,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":440,"completion_tokens_details":{"reasoning_tokens":385}},"tokens_in":440,"tokens_out":484,"duration_ms":4916,"temperature":1.0,"reasoning_tokens":385,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:55:40.040974+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reconstruct a textureless white room or a low-light corridor with known ground-truth geometry, where monocular depth and normal priors are unreliable and SfM points are sparse. If HaloGS's triangle soup and compact mesh show Chamfer distance no better than 2DGS or MVS while its rendering PSNR stays high, the claim that accurate geometry and photorealistic appearance are jointly achieved would fail: appearance could be right while geometry is wrong.","supporting_citations":[{"cited_title":"2d gaussian splatting for geometrically accurate radiance fields","cited_arxiv_id":null,"evidence_quote":"Provides the ray–splat intersection and depth/normal rendering formulation that HaloGS's triangle rasterizer builds on, plus a geometric baseline."},{"cited_title":"Finding good configurations of planar primitives in unorganized point clouds","cited_arxiv_id":null,"evidence_quote":"GoCoPP planar primitive extraction converts the triangle soup into the LoD planes and compact mesh."},{"cited_title":"Gpu-based ray-casting of quadratic surfaces","cited_arxiv_id":null,"evidence_quote":"The explicit ray–splat intersection algorithm is the geometric kernel of the custom triangle rasterizer."}],"review_version":1}