{"id":"a4775228-2cf1-42e0-9a23-023974353c7f","arxiv_id":"2506.23854","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"HiNeuS builds accurate 3D surfaces from photos by combining SDF-based visibility checks, local planar regularization, and rendering-error-weighted Eikonal constraints, reporting SOTA on several benchmarks.","lead":"HiNeuS is a neural network system that builds accurate 3D surface models from ordinary photos, even when surfaces are shiny, blank, or finely detailed. It combines three fixes for common failure modes of earlier methods and reports better accuracy on standard benchmarks, which matters for AR, robotics, and urban modeling.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 4's visibility product with β=100 and K=64 can fall below the V_j>0.9 threshold for visibly unoccluded rays because near-surface samples make σ(β f) < 1; without an excluded endpoint, Eq. 3's visible-view set may be empty.","rationale":"The reader's weakest_assumption already targets Eq. 4 vs the 0.9 threshold, and my analysis agrees; this is the most load-bearing issue because the paper's abstract, introduction, methodology, and ablations all ground the reflection-handling claim in the SDF visibility factor. The mathematical problem is not a missing error bar or an unreleased code issue; it is that, under the stated hyperparameters and sampling procedure, the mapping from 'unobstructed ray' to V_j ≈1 cannot hold, since the endpoint has f=0 and the nearest sample makes σ(β f) substantially <1. The threshold then produces a vacuous visible-view set exactly in the regime (close-up reflective objects) where the method claims its largest gains. A secondary, independent concern is that Eq. 12 uses per-point rendering error C(x), although C is defined per ray; this is also underspecified, but it is less central because the Eikonal relaxation is ablated separately and even a corrected per-point proxy may preserve the qualitative behavior. I therefore recommend no change to the CONDITIONAL verdict: the authors should release code, and the specific pass-rate test above should be run before the central claims are accepted. I do not see a reason to reject outright, because an implementation detail such as excluding a small neighbourhood of the zero crossing or annealing β would repair the mechanism, and the other two contributions plus the benchmarks may still stand.","tokens_in":15002,"tokens_out":7512,"duration_ms":87852,"concrete_test":"Obtain code or re-implement Algorithm 1 alone: take a converged SDF from a standard NeuS model on DTU scan 24, select known-visible surface points x_i^0 (e.g., points visible in 10+ cameras by mesh/COLMAP), and for each source view j compute V_j exactly as written: K=64 uniform samples over (0, D), β=100, no endpoint exclusion. Report the fraction of visible camera pairs satisfying V_j > 0.9 and the distribution of V_j. If the fraction is near zero (or if any visible view has V_j ≤0.5 when a sample lands on the zero crossing), Eq. 3 cannot define V_v as intended. Then repeat with the endpoint excluded by a margin ε = 0.1·D/K and with β annealed to 1000 to see whether any documented variant restores pass rates; this distinguishes a paper-writing bug from a genuine algorithmic failure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is Eq. 4, V_j = ∏_k σ(β f(x_k)), combined with the threshold V_j > 0.9 that defines the 'visible views' set V_v in Eq. 3. Algorithm 1 samples K=64 points uniformly in t ∈ (0, ||x_i^0 - o_j||), with the ray terminating at surface point x_i^0, where f = 0. For a genuinely visible view, f > 0 along the ray but tends to 0 at the endpoint. The closest sampled point therefore has distance to the zero crossing roughly D/(K+1) in expectation; with the stated β=100, a distance of 0.015 gives σ(β f) ≈ 0.82, and every other factor is <1. Thus V_j is systematically below 0.9 for any visible view with object-camera distance D ~ 1, and with proposal-based sampling that clusters near surfaces the product is even smaller. If no view passes the threshold, V_v is empty, λ_ambiguity is undefined/zero, and the ambiguity-aware weighting in Eq. 6/9 is inactive. The paper does not state that the endpoint is excluded, that an ε is inserted, or that β is annealed; it fixes β=100. This is not merely a numerical detail: the SDF-visibility component is one of the three headline contributions and its ablation is reported as a major source of improvement (-15.8% CD in Table 2; -1.42 dB in Table 3). As written, the mechanism is internally inconsistent; at best, the authors must be relying on an undocumented implementation detail that changes Eq. 4 or the threshold.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"HiNeuS proposes a unified neural surface reconstruction framework that combines SDF-guided visibility verification for reflective scenes, ray-aligned planar regularization for textureless regions, and rendering-error-weighted Eikonal relaxation for detail preservation. The method is evaluated on NeRF-Synthetic, GlossySynthetic, Mip-NeRF 360, and DTU, reporting state-of-the-art PSNR and Chamfer distance numbers, with ablations for each component. The central claim is that the three components jointly resolve multi-view radiance inconsistency, textureless-surface ambiguity, and detail erosion, outperforming prior reflection-aware and high-fidelity baselines.","tokens_in":15430,"tokens_out":4048,"duration_ms":48049,"significance":"If the reported results hold, the paper would represent a meaningful step forward in neural surface reconstruction, particularly for reflective and textureless scenes, and the component ablations are a useful contribution. The work is also notable for integrating visibility reasoning, local geometric regularization, and adaptive Eikonal weighting in a single optimization, rather than as separate post-processing stages. The paper ships extensive benchmark comparisons and ablation tables that are internally consistent, and no evidence of metric manipulation is apparent. However, the central visibility mechanism is currently underspecified in a way that may make it inoperative as written, and the per-point rendering error used in the Eikonal relaxation is not formally defined.","major_comments":[{"comment":"The load-bearing visibility mechanism in Eq. (4) is underspecified to the point of internal inconsistency. With K=64, beta=100, and uniform samples in (0, ||x_i^0 - o_j||), the closest sampled point to the surface zero crossing is expected to be about D/(K+1) from f=0; for an object-camera distance D on the order of 1 this gives sigma(beta f) approximately 0.82 for that point, and every other factor in the product is less than 1, so V_j cannot reach the threshold 0.9 even for an unoccluded ray. The paper does not state that the surface endpoint is excluded, that an epsilon offset is applied, or that beta is annealed; if no view satisfies V_j>0.9, then V_v is empty and lambda_ambiguity in Eq. (3) is undefined, making the weighting in Eq. (6) and Eq. (9) inactive. This directly affects the reported ablation gains attributed to SDF visibility (-15.8% CD in Table 2, -1.42 dB in Table 3), so the product formulation and threshold must be reconciled with the implementation.","section":"Section 3.5, Eq. (12)"},{"comment":"Eq. (12) defines omega(x) through ||C(x) - hat C(x)||^2, but rendering error is only defined for a ray after volume rendering in Eq. (1); no per-point color C(x) is introduced anywhere in the paper. In addition, the planarity factor lambda_pla(x) used in Eq. (12) is not connected to the discrete weights lambda_k^pla defined in Section 3.4. Please define these quantities as functions of 3D position, or state how ray-level errors are assigned to points; without this, the Eikonal relaxation in Eq. (13) cannot be implemented as written.","section":"Section 3.5, Eq. (12)"}],"minor_comments":[{"comment":"The text repeatedly claims a 21.4% Chamfer distance reduction over Ref-NeuS on GlossySynthetic, but the rounded table values (0.0038 vs 0.0048 mm) correspond to about 20.8%; please reconcile the claim with the reported precision.","section":"Abstract, Section 1, Section 5, Table 2"},{"comment":"The term 'planar-conformal regularization' is not defined; Eq. (10) enforces local SDF linearity, which is a planarity constraint but not a conformal map. Consider using a more precise name or defining the conformal sense.","section":"Section 3.4 and Section 3.5"},{"comment":"The text states that omega is proportional to 1/||C - hat C||, but Eq. (12) uses an exponential decay, not an inverse relationship; this descriptive sentence should be corrected to match the equation.","section":"Section 3.5, property (i)"},{"comment":"Section 3.4 refers to Figure 1 as an overview of the framework, but Figure 1 is a qualitative comparison image; there is no actual framework diagram, which makes the method harder to follow.","section":"Section 3.4 and Figure 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is potentially publishable if the visibility mechanism is clarified and the per-point rendering error is formally defined. The current text leaves enough ambiguity that a reader cannot reproduce the core reflection-handling component, and the inconsistency between Eq. (4), the threshold, and the stated hyperparameters is a correctness risk rather than a stylistic issue. No concerns about citation patterns or novelty disclosure beyond noting that RaNeuS is a co-author baseline, which is standard in this area."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Hi [Name],\n\nThis is a credible, well-motivated paper on neural surface reconstruction, and the headline numbers are impressive. If those numbers reproduce, it's a new state of the art on the three hard cases it targets. The unified treatment of reflections, textureless planes, and thin structure via one pipeline is a genuine combination, and the ablations are set up to show each component earns its place. The SDF-based visibility idea is a nice step beyond NeRO's learned reflections, and the planar + Eikonal relaxation pieces are sensible.\n\nBut there is a load-bearing issue that needs to be on the table. Equation 4 defines V_j as a product of sigmoids over K=64 uniform samples between the camera and the surface point. With beta=100 and a typical object distance D, the nearest sample is roughly D/(K+1) from the zero crossing, where f is about 0.015, so the sigmoid is about 0.82. That single factor already places V_j below 0.9. So the set of \"visible views\" V_v with V_j > 0.9 is likely empty for most rays, which would make lambda_ambiguity undefined. The paper never says the endpoint is excluded, or that an epsilon is inserted, or that beta is annealed. It's fixed at 100. Either the equation is missing a detail or the mechanism is internally inconsistent. Since the ambiguity weighting is their main ablation win on DTU and GlossySynthetic, this needs to be clarified before I'd trust the claimed gains.\n\nSecondary but real: Eq. 12 uses a per-point rendering error C(x), but rendering error is only defined per ray. That's sloppy notation at best. And there's no code or run-to-run variance, so it's hard to know whether the 0.43 mm DTU average is a single seed or a stable result.\n\nThe paper is worth a serious referee. The problem is fixable - exclude the endpoint, anneal beta, or change the threshold - and the overall approach is sound enough to deserve a careful look. For your group, it's a good discussion piece on what \"visibility\" means in SDF-based rendering, though I wouldn't build on the current equations until the authors respond.","headline":"Strong engineering story and numbers, but the visibility factor in Eq. 4 can't meet its own threshold as written, which needs to be resolved before the results mean what they say.","tokens_in":15879,"tokens_out":4039,"would_cite":false,"duration_ms":42519,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"HiNeuS claims a single SDF-based pipeline resolves reflection ambiguity, textureless-region drift, and Eikonal detail loss, with state-of-the-art Chamfer distances and PSNR across synthetic and real-world benchmarks.","keywords":["neural surface reconstruction","signed distance field","multi-view consistency","reflection ambiguity","textureless regions","Eikonal regularization","neural rendering","inverse rendering"],"falsifier":"Take a synthetic scene with a known thin occluder between two cameras and a target surface point. Trace the ray described in Algorithm 1, record the SDF values at every sampled point, and evaluate the product $V_j$ against the 0.9 threshold. If a sample lands at or very near the zero crossing, the sigmoid factor is about 0.5 and the product cannot exceed 0.5, so the view is classified hidden even when it is geometrically visible; repeating the same evaluation with the surface point explicitly excluded from the sample set would show whether the visibility term is a calibrated probability or an artifact of sampling choices.","tokens_in":14776,"feed_emoji":"🧊","tokens_out":10979,"duration_ms":110014,"temperature":0.7,"pith_summary":"HiNeuS is a neural surface reconstruction method that claims one training pipeline can fix three failure modes at once: reflections that break multi-view color agreement, textureless regions with almost no visual keypoints, and Eikonal smoothness penalties that sand away fine detail. The paper's central assertion is that a signed distance field (SDF) can carry all three fixes: a visibility product computed along camera rays selects which views are trustworthy for a surface point, ray-aligned planar constraints regularize low-texture areas while an appearance-based weight protects edges, and the Eikonal constraint is modulated by rendering error so geometry is allowed to sharpen where the render is still wrong. If the method works as reported, its benchmark numbers follow: average Chamfer distance of 0.43 mm on DTU, 35.00 dB PSNR on NeRF-Synthetic, and a 21.4% Chamfer-distance reduction over Ref-NeuS on specular objects. The deeper claim is that appearance and geometry constraints should evolve together during training instead of being combined as separate post-hoc losses.","feed_headline":"SDF pipeline cuts reflective 3D error by 21.4%","feed_subtitle":"Visibility checks, local planarity, and adaptive Eikonal weights reclaim low-texture, shiny surfaces, 0.43 mm on DTU.","key_machinery":"The central object is the SDF visibility factor $V_j = \\prod_{k=1}^{K} \\sigma(\\beta f(x_k))$, a product of sigmoid evaluations of the signed distance field along the ray from an auxiliary camera to a surface point, interpreted as the probability that the point is visible from that view. The same SDF supplies normals for reflected-ray tracing, the zero-crossing surface itself, and the gradient used in the planar-conformal constraint. This factor converts a hard occlusion test into a continuous, differentiable occlusion score, and its threshold $V_j > 0.9$ decides which views enter the multi-view consistency term; the paper's reflection handling, low-texture regularization, and detail preservation all hang off this one signed-distance object.","core_discovery":"The paper claims that the three known failure modes of SDF-based neural surface reconstruction are not independent, and that a single SDF can resolve all of them if the right quantities are read off it. In the paper's terms, the core discovery is that the same continuous signed distance function can serve as the surface representation, as an occlusion culler for photometric supervision, and as the carrier for adaptive geometric regularization. The visibility factor $V_j = \\prod_{k=1}^{K} \\sigma(\\beta f(x_k))$ multiplies sigmoid evaluations of the SDF along the ray from another camera to the surface point, so views blocked by intervening geometry contribute less to the color loss; the planar loss forces the SDF to be locally linear along ray-aligned patches in textureless areas while a feature-difference weight softens the constraint at edges; and the Eikonal weight $\\omega(x)$ is lowered exactly where rendering error is high, letting high-frequency geometry survive. Trained jointly with $\\mathcal{L}_{\\text{total}} = \\mathcal{L}_{\\text{rgb}} + \\mathcal{L}_{\\text{planar}} + \\mathcal{L}_{\\text{eikonal}}$, the method reports state-of-the-art Chamfer distances and PSNR across the tested synthetic and real-world benchmarks.","pith_inferences":["Editorial inference — a direct test the paper does not run: vary $\\beta$ and sample count $K$ in Eq. (4) on scenes with thin occluders; if the 0.9 threshold's behavior changes qualitatively, the visibility term is better described as a smooth occlusion heuristic than a calibrated probability.","Editorial inference — the rendering error $\\|C(x)-\\hat C(x)\\|^2$ used in Eq. (12) is naturally defined per ray, not per surface point, so its use as a spatial weight assumes some ray-to-point assignment; if that assignment is coarse, the Eikonal relaxation may be driven by nearby rather than local error.","Editorial inference — the inverse-rendering demonstrations suggest a second use for the same visibility factor: downweighting specular rays during BRDF estimation could make material decomposition more robust, a property the geometry-only benchmarks do not directly measure.","Editorial inference — the 2.32 dB PSNR gain over 3D Gaussian Splatting is a rendering metric; geometric quality against the same scenes is not reported, so a direct surface-accuracy comparison with that baseline remains open."],"forward_implications":["Reflective surfaces can be reconstructed without explicit per-surface material modeling: the SDF visibility factor decides which views enter photometric supervision, and reflected rays are traced through the same SDF to compensate indirect radiance.","Textureless regions can be regularized locally without a global smoothness penalty, because the planar loss is applied along ray-aligned patches and is softened by appearance differences at edges.","High-frequency details are no longer unconditionally sacrificed to the Eikonal constraint; where rendering error is high the constraint weight is lowered, allowing geometry to change to reduce that error.","The same training run can deliver both the best rendering PSNR and the best surface Chamfer distance on the tested benchmarks, since appearance and geometry losses are coupled rather than stacked.","The reconstructed surfaces can be reused for downstream inverse rendering, including material decomposition and relighting, because geometry and appearance are learned jointly."],"supporting_citations":[{"why":"Supplies the volume rendering integral and the radiance-field paradigm the method builds on.","marker":"[18]"},{"why":"Provides the Laplace-CDF SDF-to-density mapping that turns the signed distance field into transmittance.","marker":"[28]"},{"why":"Establishes the NeuS baseline for volume-rendered SDF surfaces that this work extends.","marker":"[23]"},{"why":"The reflection-aware baseline whose Chamfer distance the 21.4% improvement is measured against.","marker":"[6]"},{"why":"The learned-reflection-probability baseline that the paper contrasts with its SDF-based visibility formulation.","marker":"[15]"},{"why":"The high-fidelity surface baseline used for comparison in low-texture and urban scenes.","marker":"[14]"},{"why":"Supplies the NeRF-Synthetic and Mip-NeRF 360 benchmarks and baselines used for PSNR evaluation.","marker":"[1]"},{"why":"The real-world dataset used for Chamfer distance evaluation.","marker":"[10]"},{"why":"Supplies the progressive hash-encoding feature grids used in training.","marker":"[19]"}],"fun_headline_variants":["Unified SDF cuts reflective 3D error 21.4%","One SDF unifies visibility, planarity, and Eikonal for 3D","HiNeuS: single SDF recovers shiny and textureless surfaces","21.4% lower error: one SDF fixes reflective 3D flaws","Three fixes, one SDF: 21.4% lower reflective 3D error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a product of smooth on/off factors evaluated at sampled points along a camera ray genuinely measures whether a surface point is visible, and that the 0.9 cutoff cleanly separates visible from hidden views; this premise can fail if the surface point being tested lies exactly at the SDF zero crossing, where the on/off factor is 0.5.","fun_headline_variants_meta":{"raw":{"variants":["Unified SDF cuts reflective 3D error 21.4%","One SDF unifies visibility, planarity, and Eikonal for 3D","HiNeuS: single SDF recovers shiny and textureless surfaces","21.4% lower error: one SDF fixes reflective 3D flaws","Three fixes, one SDF: 21.4% lower reflective 3D error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000936,"raw_usage":{"total_tokens":4077,"prompt_tokens":1089,"completion_tokens":2988,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":705,"completion_tokens_details":{"reasoning_tokens":2881}},"tokens_in":705,"tokens_out":2988,"duration_ms":21582,"temperature":1.0,"reasoning_tokens":2881,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:30:21.638102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic scene with a known thin occluder between two cameras and a target surface point. Trace the ray described in Algorithm 1, record the SDF values at every sampled point, and evaluate the product $V_j$ against the 0.9 threshold. If a sample lands at or very near the zero crossing, the sigmoid factor is about 0.5 and the product cannot exceed 0.5, so the view is classified hidden even when it is geometrically visible; repeating the same evaluation with the surface point explicitly excluded from the sample set would show whether the visibility term is a calibrated probability or an artifact of sampling choices.","supporting_citations":[{"cited_title":"Nerf: Representing scenes as neural radiance fields for view syn- thesis","cited_arxiv_id":null,"evidence_quote":"Supplies the volume rendering integral and the radiance-field paradigm the method builds on."},{"cited_title":"V ol- ume rendering of neural implicit surfaces","cited_arxiv_id":null,"evidence_quote":"Provides the Laplace-CDF SDF-to-density mapping that turns the signed distance field into transmittance."},{"cited_title":"Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection","cited_arxiv_id":null,"evidence_quote":"The reflection-aware baseline whose Chamfer distance the 21.4% improvement is measured against."},{"cited_title":"Nero: Neural geometry and brdf reconstruction of reflective objects from multiview images","cited_arxiv_id":null,"evidence_quote":"The learned-reflection-probability baseline that the paper contrasts with its SDF-based visibility formulation."},{"cited_title":"Neuralangelo: High-fidelity neural surface reconstruction","cited_arxiv_id":null,"evidence_quote":"The high-fidelity surface baseline used for comparison in low-texture and urban scenes."},{"cited_title":"Large scale multi-view stereopsis eval- uation","cited_arxiv_id":null,"evidence_quote":"The real-world dataset used for Chamfer distance evaluation."},{"cited_title":"Instant neural graphics primitives with a mul- tiresolution hash encoding","cited_arxiv_id":null,"evidence_quote":"Supplies the progressive hash-encoding feature grids used in training."}],"review_version":1}