{"id":"d7b9c36e-a3fb-41f7-8e37-1a20656333bd","arxiv_id":"2506.13508","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A multiview geometric regularization scheme that uses MVS depth, RGB, and normal constraints to improve both geometry and rendering quality of Gaussian Splatting radiance fields.","lead":"This paper combines multiview stereo depth with Gaussian Splatting to improve the accuracy of 3D surface reconstruction from ordinary photos. The authors use MVS depth to initialize and regularize the Gaussians, yielding smoother and more precise surfaces on standard benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Eq. (3) MVS-depth gate is self-referential: it accepts MVS depths only when they agree with the current rendered median depth, so systematic MVS bias can be locked in; the geometry claim depends on this untested assumption.","rationale":"The reader's weakest assumption correctly identifies MVS reliability as the core dependency. My stress-test sharpens this: the threshold rule in Eq. (3) uses the current rendered median depth as the reference, so the filter's selectivity is coupled to the model being optimized. This makes the method vulnerable to self-confirming bias, not just random MVS noise. The paper's limitation section acknowledges MVS unreliability but does not address the feedback loop, and the chosen datasets (DTU, TnT) are ones where COLMAP MVS is strong, so the failure mode is untested. This is the most load-bearing concern because it directly conditions the geometry improvements (DTU Chamfer 0.63 vs GOF 0.74; TnT F1 0.52 vs 0.46). The concern is significant enough to warrant a concrete empirical check, but it is not fatal: the method is clearly specified, the ablations support the components, and the limitation is acknowledged in good faith. Therefore the conditional verdict stands unchanged, with the added condition that the MVS gate's behavior under biased priors be experimentally characterized.","tokens_in":16300,"tokens_out":4397,"duration_ms":47017,"concrete_test":"On DTU, where ground-truth scans exist, label each pixel by MVS depth error and by the error of the rendered depth after MVS-guided initialization. Compute the acceptance set under Eq. (3) at the start and after convergence, and quantify the fraction of accepted pixels where MVS error exceeds rendered-depth error (pixels where the loss pushes geometry away from ground truth). If this fraction is non-negligible, the filter is not selecting reliable priors. As a causal test, apply a constant 5% multiplicative bias to MVS depths in a subset of texture-rich pixels at initialization and run the full pipeline; if final Chamfer distance degrades toward the biased depths instead of recovering, the loss cannot reject systematic MVS bias.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that, after thresholding, the retained MVS depths are more accurate than the Gaussian geometry being optimized. Equation (3) implements rejection with |D_mvs - D_r| < s * D_r, where D_r is the rendered median depth from the current Gaussian model. Because D_r is the quantity being optimized, the filter is not an independent validator of MVS quality; it selects MVS depths that agree with the current geometry. If early optimization drifts (e.g., from appearance-geometry ambiguity due to specularities or lighting), D_r is biased, and the filter preferentially retains MVS depths consistent with that bias while discarding corrective depths. The MVS-guided initialization mitigates but inherits the same bias when MVS is systematically wrong. The limitation section acknowledges unreliable MVS but not this feedback coupling. On DTU and Tanks and Temples, COLMAP MVS is strong, so this failure mode may not appear; no experiment isolates it. The improvement over GOF is consistent per scene, but those scenes do not stress the assumption where MVS is biased in well-textured regions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a multiview geometric regularization strategy for Gaussian Splatting, building on Gaussian Opacity Fields (GOF). The method has three main components: an MVS-guided initialization that produces a dense, geometrically informed starting point; a median-depth-based multiview relative depth loss with a rendering-uncertainty weight that aligns Gaussian positions with MVS depth estimates; and multiview extensions of RGB photometric, normal-consistency, and depth-distortion losses. Experiments on DTU, Tanks and Temples, and Mip-NeRF360 report improved surface reconstruction over GOF (DTU mean Chamfer distance 0.63 vs 0.74 mm; TnT mean F1 0.52 vs 0.46) with roughly maintained or slightly improved novel-view synthesis metrics. The paper includes component-wise ablations on one Mip-NeRF360 scene and aggregate ablations on DTU/TnT.","tokens_in":16484,"tokens_out":4341,"duration_ms":42470,"significance":"If the reported results hold, the paper makes a useful contribution to explicit Gaussian-splatting surface reconstruction by demonstrating that MVS depth priors can be combined with splatting optimization to improve geometry without sacrificing rendering quality. The paper's strengths are its clear component-wise ablation study (Figures 4 and 5), the consistent per-scene improvement over GOF on both DTU and TnT, and the breadth of comparisons against implicit and explicit baselines. The proposed relative depth loss with median-depth rendering and accumulated-alpha uncertainty is simple and plausible. However, the central claim depends on an assumption, stated in Section 7, that MVS depths are more accurate than the unregularized Gaussian geometry; the paper does not experimentally stress this assumption, and the rejection rule in Eq. (3) has a feedback-coupling property that is not analyzed. These concerns are addressable with additional experiments.","major_comments":[{"comment":"The rejection gate in Eq. (3) uses the rendered median depth D_r as the reference for accepting or rejecting MVS depths, but D_r is the very quantity being optimized by this loss. This creates a feedback coupling: if the early optimization drifts (for instance, due to appearance-geometry ambiguity in specular or view-dependent regions), the gate preferentially retains MVS depths consistent with the biased D_r and discards corrective depths, so a systematic MVS bias can be locked in rather than corrected. The Limitations section (Section 7) acknowledges that MVS may be unreliable, but it does not address this coupling. Because the central geometry claim relies on the retained MVS depths being more accurate than the current Gaussian geometry, please add an experiment that directly tests this assumption, for example by injecting a known systematic bias into the MVS depths on a synthetic scene with ground truth and measuring whether the filter recovers or amplifies the bias.","section":"Section 5.1, Eq. (3)"},{"comment":"The main geometry comparisons report single-run numbers without error bars, significance tests, or multiple seeds. The per-scene improvements over GOF are often small (e.g., DTU scan24: 0.43 vs 0.50 mm; TnT Ignatius: 0.67 vs 0.68 F1), so without variance information it is difficult to determine whether the mean improvement from 0.74 to 0.63 mm (DTU) and from 0.46 to 0.52 F1 (TnT) is robust to optimization noise. Please report variance across at least three seeds for the primary comparison with GOF, or justify why single runs are representative.","section":"Section 6.3, Tables 1 and 2"},{"comment":"The paper claims that the uncertainty term U = Σ_i ω_i is a key component of the relative depth loss, but no ablation isolates its contribution. Figures 4 and 5 vary multiple components simultaneously (for example, (e) vs (f) adds both MVS-guided initialization and multiview normal/depth-distortion losses together with the uncertainty-weighted depth loss), so the reader cannot tell whether U improves results over an unweighted relative depth loss. Please add an ablation with U removed or set to 1, and report its effect on both DTU and TnT metrics.","section":"Section 5.1 and Figures 4/5"}],"minor_comments":[{"comment":"The introduction and abstract claim that the method outperforms state-of-the-art models in both novel-view synthesis and surface reconstruction, but Table 3 shows that 3DGS-MCMC achieves higher PSNR, SSIM, and lower LPIPS on Mip-NeRF360, and the paper later states it is second-best overall. Please rephrase the NVS claim to match the reported results.","section":"Section 1, Abstract"},{"comment":"The threshold parameter s is used in Eq. (3) but is only defined later in Section 6.1; please define it near the equation or refer forward explicitly.","section":"Section 5.1, Eq. (3)"},{"comment":"Reference [LHH25] is cited for using monocular depth priors to handle weakly textured regions, but the reference title 'Multi-plane Gaussian Splatting for compact scenes rendering' does not appear to match that description; please verify the citation.","section":"Section 2.1 referenced [LHH25]"},{"comment":"The training-time notation '58.1m + α' and '75.1m + α' with α defined as approximately 4m (DTU) and 35m (TnT) is confusing because α has different meanings in the two tables; consider reporting a single total time or a clearer breakdown.","section":"Tables 1 and 2"},{"comment":"The heading 'Neural Radience Field' contains a typo; it should read 'Neural Radiance Field'.","section":"Section 2.1 heading"},{"comment":"The caption uses 'multview normal and depth distortion losses' in the description of configuration (f); this should be 'multiview'.","section":"Figure 4 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of Computer Graphics Forum and the core idea is sensible. The main concern is the untested feedback coupling in the depth-filtering rule; an experiment with a known biased depth prior would substantially strengthen the paper. The lack of error bars is also a standard expectation for empirical claims of this type. I would encourage the authors to address these points before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a solid methods paper on a narrow but active problem: pulling accurate surfaces out of Gaussian Splatting. The new parts are a median-depth-based multiview relative depth loss with uncertainty, an MVS-guided initialization that seeds Gaussians from filtered depth maps, and multiview versions of the usual normal/depth-distortion regularizers. The ablations are careful and each component earns its place. The headline numbers are real: DTU Chamfer 0.63 versus 0.74 for GOF, TnT F1 0.52 versus 0.46, with NVS quality essentially preserved. For explicit GS surface reconstruction, that is a meaningful step, not a new paradigm.\n\nWhat I like: the paper is honest about its central assumption. The Limitations section states plainly that MVS depths must be more accurate than unregularized GS geometry, especially in textured regions. That is the load-bearing condition, and the authors do not hide it. The experiments on DTU and TnT are consistent across scenes, and the ablation in Figures 4 and 5 shows that each proposed piece contributes.\n\nSoft spots, in order of seriousness. First, no code or data is released, and there are no error bars or significance tests. The mean Chamfer gains are consistent, but we are taking the authors' word for the exact numbers. Second, the introduction overstates the NVS result: Table 3 shows parity with GOF, not an improvement, for PSNR and SSIM on Mip-NeRF360; the improvement is mainly in LPIPS. That mismatch should be fixed. Third, the stress-test concern about Eq. (3) has some teeth: the gate that decides which MVS depths to keep uses the current rendered depth as reference, so it is not an independent validator. If the optimization drifts early and MVS bias happens to agree with that drift, the filter will retain the wrong depths. The MVS-guided initialization mitigates this by starting from MVS points, but it inherits the same bias when MVS is systematically wrong. The paper never isolates this failure mode, and the benchmarks used (COLMAP on DTU/TnT) are exactly where MVS is strongest. I would not call this fatal — the authors explicitly acknowledge the assumption — but it is the main empirical gap.\n\nBottom line: this is a well-executed, well-ablated paper that deserves a serious referee. If it were submitted to me, I would send it out rather than desk reject. The claim that matters — MVS priors improve GS geometry — is supported on the tested benchmarks, with the caveat that the failure mode where MVS is biased is not explored. Worth citing for anyone working on GS-based surface reconstruction.","headline":"A well-ablated, honest methods paper that shows real geometry gains from MVS-guided regularization of Gaussian Splatting, with the main caveat that the MVS-bias failure mode is untested.","tokens_in":17101,"tokens_out":2725,"would_cite":true,"duration_ms":24862,"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":"Gaussian Splatting surfaces become markedly more accurate when MVS depth priors guide both initialization and a multiview median-depth loss.","keywords":["Gaussian Splatting","multiview stereo","surface reconstruction","median depth loss","radiance fields","geometric regularization","MVS-guided initialization"],"falsifier":"A concrete test: take a scene with a ground-truth mesh, perturb all MVS depths by a constant fraction smaller than the annealing threshold (say 4 percent) and rerun the pipeline; if the output surface shifts by roughly that amount, the method is letting systematically biased priors through the gate.","tokens_in":16024,"feed_emoji":"📐","tokens_out":6582,"duration_ms":56886,"temperature":0.7,"pith_summary":"3D Gaussian Splatting renders fast but its reconstructed surfaces are often noisy or wrong. This paper claims that fusing multiview stereo (MVS) depth with Gaussian Splatting at two points—initialization and optimization—produces smooth, accurate geometry while keeping rendering quality. The key move is a relative depth loss that compares the rendered median depth against MVS depth, only activates when the two roughly agree, and weights itself by how certain the rendering is. On the DTU benchmark the mean Chamfer distance falls from 0.74 (GOF baseline) to 0.63, and on Tanks and Temples the F1 score rises from 0.46 to 0.52. Those numbers make the claim concrete: explicit splatting methods no longer have to choose between speed and geometric fidelity.","feed_headline":"MVS depth priors sharpen Gaussian splatting geometry","feed_subtitle":"A median-depth loss plus MVS-guided starts beats prior explicit surface methods on DTU and Tanks and Temples.","key_machinery":"The load-bearing component is the multiview median-depth relative loss $L_{\\mathrm{rel}} = \\lvert 1 - D_{\\mathrm{mvs}}/D_r \\rvert \\cdot U \\cdot \\mathbb{1}(\\lvert D_r - D_{\\mathrm{mvs}}\\rvert < s D_r)$, where $D_r$ is the rendered median depth (the largest depth with accumulated transmittance above 0.5), $U$ is the accumulated opacity along the ray, and $s$ is a threshold annealed from 0.15 to 0.05. This loss converts MVS depths into a soft attractor that is active only when the current Gaussian-splatting depth and the MVS depth roughly agree, and its pull is scaled by rendering certainty. It is evaluated across triplets of adjacent views, together with multiview extensions of RGB, normal-consistency, and depth-distortion losses. The companion mechanism is MVS-guided initialization: voxel-filtered MVS points seed the Gaussians, and an initial phase that freezes positions while pruning low-opacity Gaussians removes outliers before the main optimization.","core_discovery":"The paper's central claim is that MVS-derived depth and Gaussian Splatting's optimized points are complementary, and that an optimization scheme can harvest both. MVS patch matching with epipolar constraints is reliable in well-textured regions but noisy near boundaries; Gaussian Splatting handles boundaries but can drift under view-dependent appearance. The authors exploit this by initializing Gaussians from filtered MVS point clouds, then adding a multiview relative depth loss whose reference is the rendered median depth, with accumulated opacity as uncertainty. The loss rejects MVS depths that deviate beyond an annealed threshold and pulls the Gaussians toward the remaining depths. Ablations show that neither initialization nor the depth loss alone is enough: the full combination is what removes noise and keeps surfaces smooth.","pith_inferences":["A direct extension would be to replace the hard indicator threshold with a robust weighting, which could recover scenes where MVS depths are systematically biased by less than the threshold.","The same median-depth-plus-confidence recipe may transfer to other explicit primitives, such as surfels or 2D disks, that already have ray-splat intersection logic.","Because the method depends on MVS patch matching, its gains plausibly shrink on textureless or heavily specular scenes; that is a testable boundary rather than a contradiction.","One could isolate the value of epipolar consistency by feeding monocular depth through the identical loss and initialization; if the gains persist, the essential ingredient is the prior, not the multiview matching."],"forward_implications":["On DTU, the mean Chamfer distance drops from 0.74 for GOF to 0.63, and on Tanks and Temples the mean F1 score rises from 0.46 to 0.52.","On Mip-NeRF360, novel-view synthesis stays close to the strongest appearance-only methods, with LPIPS 0.197 on outdoor and 0.172 on indoor scenes.","The additional cost is roughly 4 minutes of MVS depth estimation on DTU and 35 minutes on Tanks and Temples, plus longer optimization from denser Gaussian sets.","The method remains an explicit splatting pipeline, so it keeps the fast, differentiable rendering of GOF while improving geometric fidelity."],"supporting_citations":[{"why":"Supplies the PatchMatch multiview stereo depth maps that serve as the geometric prior and initialization source.","marker":"[SZPF16]"},{"why":"The Gaussian Opacity Fields baseline whose rendering formulation and loss the method extends and improves upon.","marker":"[YSG24]"},{"why":"Source of the single-view normal-consistency and depth-distortion regularizers and the median-depth convention that the multiview losses build on.","marker":"[HYC*24]"},{"why":"The original 3D Gaussian Splatting representation and optimization framework that GOF and this method inherit.","marker":"[KKLD23]"},{"why":"Tanks and Temples benchmark used to evaluate large-scale surface reconstruction F1 scores.","marker":"[KPZK17]"},{"why":"DTU benchmark used to evaluate object-scale Chamfer distances.","marker":"[JDV*14]"},{"why":"Mip-NeRF360 dataset and evaluation protocol used for novel-view synthesis comparison.","marker":"[BMV*22]"},{"why":"TSDF fusion used to extract meshes from rendered median depth maps on DTU.","marker":"[CL96]"}],"fun_headline_variants":["MVS + Gaussian splatting: complementary geometry pairing","Multiview depth constraints sharpen Gaussian splatting","Median-depth loss refines Gaussian splatting geometry","MVS-guided initialization improves Gaussian splatting","Combining MVS and Gaussian splatting for accurate radiance fields"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that MVS depth estimates, after heuristic geometric filtering, are accurate enough in textured regions to serve as a valid geometric prior, and that the threshold rule $|D_r - D_{\\mathrm{mvs}}| < s D_r$ reliably separates good MVS depths from bad ones.","fun_headline_variants_meta":{"raw":{"variants":["MVS + Gaussian splatting: complementary geometry pairing","Multiview depth constraints sharpen Gaussian splatting","Median-depth loss refines Gaussian splatting geometry","MVS-guided initialization improves Gaussian splatting","Combining MVS and Gaussian splatting for accurate radiance fields"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000889,"raw_usage":{"total_tokens":3829,"prompt_tokens":932,"completion_tokens":2897,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2817}},"tokens_in":548,"tokens_out":2897,"duration_ms":19631,"temperature":1.0,"reasoning_tokens":2817,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:59:17.966686+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test: take a scene with a ground-truth mesh, perturb all MVS depths by a constant fraction smaller than the annealing threshold (say 4 percent) and rerun the pipeline; if the output surface shifts by roughly that amount, the method is letting systematically biased priors through the gate.","supporting_citations":[],"review_version":2}