{"id":"7a6b1731-79a5-4b1c-b4bb-8dc1815f2999","arxiv_id":"2507.15602","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"SurfaceSplat combines SDF-based coarse meshes with Gaussian splatting to improve sparse-view reconstruction and rendering, but the ablations do not isolate the effect of each component.","lead":"This paper presents a hybrid 3D reconstruction pipeline that uses a signed distance field to seed a Gaussian splatting model and then uses that model's rendered images to refine the surface. The method improves surface reconstruction and novel-view rendering on two datasets under sparse-view conditions, but the contribution of each component is not cleanly isolated.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pseudo-view expansion is confounded with normal loss and mesh cleaning and may be circular: ablation baselines are plain Voxurf/GOF, so the central 3DGS-for-SDF gain is not isolated.","rationale":"The paper presents a clear, coherent pipeline and reports large empirical gains on MobileBrick and DTU (Tables 1-2), which is real evidence that the combined system works better than the prior methods it compares against. The stated central claim, however, is specifically that the two-way coupling—SDF-initialized 3DGS rendering pseudo-views that then refine the SDF—is what produces those gains. For that claim to hold, the pseudo-view expansion must provide supervision that is both new and reliable. The weakest point is that the ablation design does not isolate this mechanism: the 'baseline' rows in the key ablations (Tables 3, 5, and 9) are numerically identical to plain Voxurf and GOF, not to SurfaceSplat with a single component removed. Hence the measured improvements could be driven by the added normal loss, mesh cleaning, or the SDF-mesh initialization of 3DGS, all of which are also novel relative to Voxurf. The second concern is epistemic rather than statistical: because the pseudo-view renderer is initialized from the same coarse SDF mesh and trained on the same original images, rendered novel views are not an independent source of geometric truth. They can still be useful as a learned interpolant, but the paper provides no test showing that the SDF is not simply fitting its own earlier output. These concerns do not imply the method is invalid; the benchmark gains may be genuine. Rather, they show that the central attribution is not yet demonstrated. A four-arm ablation, as specified, would separate the pseudo-view contribution from the other components and would also reveal whether COLMAP-initialized 3DGS pseudo-views give the same benefit, thereby addressing circularity. The absence of quantitative DTU rendering results and of error bars are additional support gaps, but they are secondary to the mechanism-isolation problem. The appropriate verdict remains conditional: accept once the pseudo-view expansion is cleanly isolated with a matched baseline and the circularity question is resolved.","tokens_in":1192,"tokens_out":1111,"duration_ms":96590,"concrete_test":"Run a controlled ablation on a fixed subset of MobileBrick/DTU scenes with four arms: (A) SurfaceSplat's SDF stage with normal loss (Eq. 4), mesh cleaning (Eqs. 5-6), and matched fine-stage training, but no expanded pseudo-views; (B) A plus pseudo-views rendered by the same SDF-initialized 3DGS; (C) A plus pseudo-views rendered by a 3DGS initialized from COLMAP points only; (D) A plus an equal number of real held-out views. Report F1, CD, and PSNR. If B does not beat A by the margins implied by Tables 3/5, or if C matches B, the pseudo-view contribution is either absent or circular; if D greatly exceeds B, the pseudo-views are a weak oracle.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that pseudo-views rendered by 3DGS provide new supervision that refines SDF detail (Sec. 3.3, Eq. 7). The evidence for this mechanism is not isolated. In Table 5 and Table 9, the rows labeled 'Baseline' or 'SDF-based method w/o 3DGS' are not SurfaceSplat minus pseudo-views; they are unmodified Voxurf. The numbers coincide exactly: Aston F1=55.8 appears in both Table 5 and Table 7 (Voxurf), DTU scan69 CD=2.94 appears in both Table 3 and Table 2 (Voxurf), and Table 9's F1=62.42 / PSNR=18.34 match Table 1's Voxurf row. Likewise, '3DGS-based method w/o SDF' matches GOF. Therefore the reported improvement could come entirely from SurfaceSplat's added monocular normal loss (Eq. 4), mesh cleaning (Sec. 3.2), or SDF-mesh initialization of 3DGS, rather than from the rendered pseudo-views. Compounding this, the pseudo-views are rendered by a 3DGS initialized from the same coarse SDF and trained on the same images, so they may encode the initial SDF's geometry instead of independent evidence; Eq. 7 is deterministic splatting of G, and no confidence or consistency filter is applied before adding {I_new} to the SDF training set. Thus the two-way coupling at the heart of the paper is underdetermined by the present ablations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SurfaceSplat, a sparse-view reconstruction pipeline that alternates between SDF-based surface reconstruction and 3D Gaussian Splatting. A coarse SDF mesh (Voxurf) is cleaned, sampled to initialize 3DGS, and then 3DGS renders additional novel-view images that are added to the training set to refine the SDF at higher voxel resolution. Experiments on DTU and MobileBrick report state-of-the-art Chamfer Distance/F1 and PSNR/PSNR-F, with ablations on pose expansion, number of rendered views, and 3DGS initialization.","tokens_in":16679,"tokens_out":5092,"duration_ms":47721,"significance":"If the proposed two-way coupling is the cause of the reported gains, the paper is a solid step toward combining dense and sparse scene representations for sparse-view reconstruction. The manuscript is clearly written, the datasets are standard, the code is promised for release, and the experimental breadth (two datasets, multiple sparsity levels, per-scene tables, and a SparseCraft variant) is a genuine strength. The main reservation is that the ablations do not isolate the pseudo-view-expansion mechanism from the other modifications (normal loss, mesh cleaning, SDF-initialized 3DGS), and the pseudo-views may inherit the coarse SDF geometry rather than provide independent supervision; these points must be addressed before the central claim is fully supported.","major_comments":[{"comment":"The ablations do not isolate the contribution of the 3DGS-rendered pseudo-views. In Table 9, the row 'SDF-based method w/o 3DGS' reports F1=62.42, CD=13.3, PSNR=14.34, PSNR-F=18.34, which are exactly the Voxurf numbers in Table 1; in Table 5, the 'Baseline' F1=55.8 is exactly the Voxurf value for aston in Table 7; and in Table 3, the 'Baseline' on DTU scan69 is CD=2.940, again the Voxurf value. Thus these baselines are plain Voxurf, not SurfaceSplat with only the pseudo-view expansion removed. The reported 3DGS-for-SDF gain (about 5 F1 points in Table 5 and 6.5 F1 points in Table 9) could therefore be caused entirely by the monocular normal loss (Eq. 4), the mesh cleaning (Sec. 3.2), the SDF-mesh initialization of 3DGS, or some combination, rather than by the rendered pseudo-views that are the mechanism claimed in Sec. 3.3 and the abstract. I ask for an ablation that removes only the pseudo-views from the pipeline (keeping the normal loss, cleaning, and 3DGS stage) and for the baseline rows to be relabeled as Voxurf if that is what they are.","section":"Sec. 4.3, Tables 3, 5, 9"},{"comment":"The pseudo-views are rendered by a 3DGS model initialized from the same coarse SDF mesh and trained on the same input images, so the supervision they provide to the SDF may be largely circular: the SDF is refined on images that already encode its own coarse geometry. Equation (7) is a deterministic splatting of G with no confidence or multi-view-consistency filter before {I_new} is added to the SDF training set. I do not claim the mechanism is certainly circular, but the present experiments do not rule it out. A concrete and feasible test would be to compare the current pseudo-view expansion against expansion using views rendered from a 3DGS initialized only from COLMAP points, and to measure whether the refinement improves on held-out real views; alternatively, report the photometric or geometric consistency error between rendered pseudo-views and the original inputs for the added viewpoints.","section":"Sec. 3.3, Eq. (7)"},{"comment":"The improvements reported in Tables 4 and 5 are consistent with an alternative explanation: the monocular normal loss alone, plus the largest-component mesh cleaning, may be responsible for most of the gains, and the pseudo-view expansion may add little once those components are present. Table 4 shows that 3DGS rendering improves from 19.13 to 20.45 with 'normal and mesh clean', but it does not include a row with the normal loss alone or with mesh cleaning alone, so the individual contributions are not separable. Because the normal loss is already a strong prior in sparse-view SDF reconstruction (as in MonoSDF), the paper should show a row of SurfaceSplat without the rendered views but with all other components, or equivalently a row with pseudo-views but without the normal loss, to attribute the Table 3 and Table 9 gains specifically to the 3DGS-for-SDF coupling.","section":"Sec. 3.2, Eq. (4) vs. Sec. 3.3"}],"minor_comments":[{"comment":"The caption contains a typo: 'pose expansion strategies for in MobileBrick' should read 'pose expansion strategies on MobileBrick'.","section":"Sec. 4.3, Table 5 caption"},{"comment":"The heading 'Combing 3DGS and SDF' should be 'Combining 3DGS and SDF'.","section":"Sec. 2.3 heading"},{"comment":"The COLMAP reference is given as [33] in Sec. 3.2 and [31] in Sec. 4.1; the citation index should be consistent.","section":"Sec. 3.2 and Sec. 4.1 references"},{"comment":"The symbol R is used for the rendering operator in Eq. (10) and for camera rotation matrices elsewhere; please use distinct notation to avoid ambiguity.","section":"Sec. 3.4, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is promising but the central mechanism is not yet isolated. The reviewer's conditional assessment is fair, and the identical numbers between the ablations and the plain Voxurf rows are a concrete issue that should be fixed by re-running the appropriate ablations. I would also encourage the editor to request a clear statement in the ablation tables of exactly which components are included in each baseline row. The paper may also benefit from a comparison against a 3DGS-to-SDF refinement that uses pseudo-views generated without SDF-initialized 3DGS, to address the circularity concern."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a practical and clearly-written hybrid for sparse-view reconstruction, and the reported gains on DTU and MobileBrick are worth paying attention to. But the paper's signature mechanism—3DGS-rendered pseudo-views refining the SDF—is not isolated in the experiments, and the ablations as presented don't support the claim that the rendered views are what helps.\n\nWhat's new: the specific two-stage pipeline (SDF coarse mesh -> 3DGS initialization -> 3DGS-rendered views as extra training data for SDF) is a combination I haven't seen in prior joint optimization work like NeuSG, GSDF, or 3DGSR, which target dense views and don't use rendered views to augment SDF training. The normal loss and mesh cleaning for the 3DGS initialization are sensible, and Table 4 shows they help rendering (foreground PSNR from 19.13 to 20.45). The full method beats strong baselines by a wide margin (DTU 5-view CD 1.45 vs MonoSDF 1.70; MobileBrick F1 68.97 vs Voxurf 62.42). That's a real result.\n\nThe soft spot: the ablation evidence for '3DGS for Enhanced SDF' is confounded. In Table 5 and Table 9, the 'Baseline' or 'SDF-based method w/o 3DGS' rows are simply Voxurf, not SurfaceSplat minus the pseudo-view expansion. This matches the stress-test note: the numbers coincide exactly with the Voxurf row in Table 1. So the +5 F1 points could come entirely from the monocular normal loss (Eq. 4), the mesh cleaning, or the better 3DGS initialization—not from the rendered views. The paper needs a proper ablative: Voxurf + normal loss, then add pseudo-views, and show the increment. The circularity concern is real too—the pseudo-views come from a 3DGS initialized from the same coarse SDF and trained on the same images, and Eq. 7 has no confidence or consistency filter. It's not necessarily fatal, but the current experiments don't prove the pseudo-views are adding independent information.\n\nMinor issues: no error bars, DTU rendering results are only qualitative, and code is promised but not yet released. All fixable. The authors do honestly acknowledge limitations (object-level scenes, struggles with two images).\n\nOverall, this paper deserves a serious referee. The method is practical, the results are strong, and the narrative is coherent. But reviewers should push for a cleaner ablation that isolates the pseudo-view contribution, and ideally for code release. I would not cite it for the mechanism as currently evidenced, but I would cite the empirical result if the authors supply the missing breakdowns.","headline":"Practical two-stage hybrid for sparse-view reconstruction with strong benchmark gains, but the pseudo-view mechanism is not isolated in the ablations and the current evidence is confounded.","tokens_in":17255,"tokens_out":2353,"would_cite":true,"duration_ms":23356,"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":"SurfaceSplat claims that a two-way SDF–Gaussian loop—coarse mesh seeding splats, rendered views refining the mesh—makes sparse-view reconstruction and rendering both improve.","keywords":["sparse-view reconstruction","signed distance function","3D Gaussian splatting","novel view synthesis","surface reconstruction","point cloud initialization","pseudo-view expansion","hybrid 3D representation"],"falsifier":"Render the expansion views with a Gaussian model initialized from structure-from-motion points instead of from the SDF mesh, keeping everything else identical; if the reported F1 and Chamfer-distance gains over the no-expansion baseline vanish or fall to the level of adding only the monocular normal loss, the pseudo-views are not independent supervision.","tokens_in":16091,"feed_emoji":"🔁","tokens_out":11010,"duration_ms":108869,"temperature":0.7,"pith_summary":"Sparse-view 3D reconstruction fails in two complementary ways: dense signed-distance-function (SDF) models capture global shape but smooth over fine detail, while 3D Gaussian splatting keeps detail but loses global coherence. This paper tries to establish that running the two representations in a loop fixes both weaknesses. A coarse SDF mesh is cleaned and converted into point-cloud initialization for the Gaussians, and the trained Gaussians render extra camera viewpoints that are fed back as training images for a finer SDF pass. On the DTU and MobileBrick benchmarks the paper reports better mesh metrics and novel-view PSNR than the compared SDF-only, Gaussian-only, and joint baselines. The practical stake is improving sparse-view reconstruction with geometric initialization and pseudo-view expansion, without large pretrained generative models.","feed_headline":"Two-way SDF-Gaussian loop wins sparse-view reconstruction","feed_subtitle":"A coarse mesh seeds the splats; their rendered pseudo-views refine the SDF, improving DTU and MobileBrick.","key_machinery":"The load-bearing object is the two-way data flow between an SDF voxel grid and a set of 3D Gaussians. The SDF-to-Gaussian link is a depth-based surface-point sampler: render depth maps of the coarse mesh at the training views, sample visible pixels, and unproject them to colored 3D points that replace the typical sparse structure-from-motion points as the geometric prior. The Gaussian-to-SDF link is a pose-expansion renderer: cubic-spline interpolation or small position perturbation generates new camera poses, the Gaussians render images from those poses, and the rendered images join the real ones as supervision for a finer SDF pass. Two smaller mechanisms keep the loop stable: a monocular-normal consistency loss that suppresses floaters in the coarse mesh, and a connected-component filter that removes non-dominant mesh pieces before sampling.","core_discovery":"The central claim, stated on the paper's own terms, is that a single optimization cycle connecting SDF and 3D Gaussian splatting outperforms either representation on its own for sparse inputs. In the SDF-to-Gaussian direction, marching cubes extracts a coarse mesh; connected-component cleaning and a monocular-normal consistency loss remove floating artifacts; depth maps of the cleaned mesh are unprojected to sample visible surface points that initialize the Gaussians, combined with sparse structure-from-motion points for background. In the Gaussian-to-SDF direction, the trained Gaussians render images from perturbed or spline-interpolated camera poses, and those rendered images are added to the original inputs to train a higher-resolution SDF grid. The paper reports that this loop lifts foreground PSNR for rendering and F1/Chamfer distance for geometry on both MobileBrick and DTU, and that a second cycle gives only marginal gains, so the default configuration is one cycle.","pith_inferences":["If the two-way loop is the active ingredient, the same recipe should transfer to casually captured video with irregular pose coverage, where uniform sparse sampling is unavailable and structure-from-motion points are unreliable; a test would be to run the pipeline on a handheld 360-degree video with pose estimates from a SLAM system.","The near-flat gain from a second cycle (F1 68.97 to 69.14, PSNR-F 20.45 to 20.55) suggests that after one expansion the bottleneck shifts from view count to the SDF model's capacity or the quality of the rendered pseudo-views, not to more expansion.","Because the Gaussian model is initialized from the same SDF mesh and trained on the same images, the rendered views may be biased toward the initial coarse geometry; we would test this by replacing the expansion images with views rendered by a Gaussian model initialized purely from structure-from-motion points and checking whether the SDF refinement gain survives.","The ablation design leaves open how much of the gain comes from the monocular normal loss rather than from the pseudo-views; a controlled experiment adding the normal loss to the no-expansion baseline would separate the two contributions."],"forward_implications":["If the central claim holds, sparse-view reconstruction can be improved by better geometric initialization alone, without generative priors: the SDF mesh supplies global structure and the Gaussians supply high-frequency detail.","The rendered pseudo-views are the active data-expansion mechanism: adding ten rendered views to ten real inputs increases the reported F1 score on MobileBrick, with gains such as +9.61 at five input views and +2.71 at ten input views, and reduces DTU Chamfer distance.","Mesh-sampled initialization benefits 3DGS rendering directly: foreground PSNR on MobileBrick rises from 19.13 with structure-from-motion points to 20.45 with the cleaned, normal-supervised mesh sampling.","The loop is representation-agnostic rather than tied to one SDF method: replacing the voxel-grid SDF with another surface reconstruction backbone in the supplementary experiments retains the reported DTU improvement."],"supporting_citations":[{"why":"Supplies the voxel-grid SDF reconstruction backbone whose coarse stage, losses, and fine stage are reused as the SDF half of the loop.","marker":"[43]"},{"why":"Supplies the Gaussian splatting model that is initialized from mesh-sampled points and renders the novel-view expansion images.","marker":"[17]"},{"why":"Supplies the monocular surface-normal predictions used by the normal consistency loss that cleans the coarse mesh before sampling.","marker":"[51]"},{"why":"Supplies the MobileBrick dataset and its F1/Chamfer/PSNR evaluation protocol for the main sparse-view comparisons.","marker":"[19]"},{"why":"Supplies the DTU dataset and Chamfer-distance benchmark used for the 5-view and 3-view reconstruction comparisons.","marker":"[16]"},{"why":"Supplies the sparse structure-from-motion point cloud that is combined with the mesh-sampled points and serves as the standard initialization compared against.","marker":"[33]"},{"why":"A Gaussian-based surface reconstruction baseline the paper uses to establish the rendering-versus-geometry gap that the hybrid loop closes.","marker":"[57]"},{"why":"Defines the SDF volume-rendering formulation used by the SDF stage.","marker":"[38]"}],"fun_headline_variants":["SDF-Gaussian loop sharpens sparse-view 3D","Two-way splat-SDF loop beats sparse-view baselines","Coarse SDF seeds splats, splats refine SDF","SurfaceSplat fuses SDF and splats for better 3D","One loop links SDF and splats for sparse views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The premise that carries the method is that the Gaussian-rendered novel views provide genuinely new, reliable supervision for the SDF, even though the Gaussian model was initialized from the same SDF mesh and trained on the same photographs.","fun_headline_variants_meta":{"raw":{"variants":["SDF-Gaussian loop sharpens sparse-view 3D","Two-way splat-SDF loop beats sparse-view baselines","Coarse SDF seeds splats, splats refine SDF","SurfaceSplat fuses SDF and splats for better 3D","One loop links SDF and splats for sparse views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000764,"raw_usage":{"total_tokens":3343,"prompt_tokens":853,"completion_tokens":2490,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":469,"completion_tokens_details":{"reasoning_tokens":2401}},"tokens_in":469,"tokens_out":2490,"duration_ms":18917,"temperature":1.0,"reasoning_tokens":2401,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:27:07.782892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Render the expansion views with a Gaussian model initialized from structure-from-motion points instead of from the SDF mesh, keeping everything else identical; if the reported F1 and Chamfer-distance gains over the no-expansion baseline vanish or fall to the level of adding only the monocular normal loss, the pseudo-views are not independent supervision.","supporting_citations":[{"cited_title":"V oxurf: V oxel-based efficient and accurate neural surface reconstruction","cited_arxiv_id":null,"evidence_quote":"Supplies the voxel-grid SDF reconstruction backbone whose coarse stage, losses, and fine stage are reused as the SDF half of the loop."},{"cited_title":"3d gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian splatting model that is initialized from mesh-sampled points and renders the novel-view expansion images."},{"cited_title":"Metric3d: Towards zero-shot metric 3d prediction from a single image","cited_arxiv_id":null,"evidence_quote":"Supplies the monocular surface-normal predictions used by the normal consistency loss that cleans the coarse mesh before sampling."},{"cited_title":"Mobilebrick: Building lego for 3d reconstruction on mobile devices","cited_arxiv_id":null,"evidence_quote":"Supplies the MobileBrick dataset and its F1/Chamfer/PSNR evaluation protocol for the main sparse-view comparisons."},{"cited_title":"Large scale multi-view stereopsis evalu- ation","cited_arxiv_id":null,"evidence_quote":"Supplies the DTU dataset and Chamfer-distance benchmark used for the 5-view and 3-view reconstruction comparisons."},{"cited_title":"Photo tourism: exploring photo collections in 3d","cited_arxiv_id":null,"evidence_quote":"Supplies the sparse structure-from-motion point cloud that is combined with the mesh-sampled points and serves as the standard initialization compared against."},{"cited_title":"Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes","cited_arxiv_id":null,"evidence_quote":"A Gaussian-based surface reconstruction baseline the paper uses to establish the rendering-versus-geometry gap that the hybrid loop closes."},{"cited_title":"Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction","cited_arxiv_id":null,"evidence_quote":"Defines the SDF volume-rendering formulation used by the SDF stage."}],"review_version":1}