{"id":"37bbd246-9f38-471e-8014-0a1f0275aedf","arxiv_id":"2412.14939","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GURecon distills multi-view photometric consistency into a continuous 3D geometric uncertainty field for neural surfaces, improving uncertainty estimation and incremental reconstruction.","lead":"GURecon learns a 3D geometric uncertainty field for neural surface reconstructions using multi-view patch consistency as supervision, with no ground-truth geometry required. It reports better alignment between predicted uncertainty and true reconstruction error than several NeRF-based baselines, and uses the uncertainty to improve next-best-view selection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The multi-view consistency pseudo-label is not a reliable proxy for geometric error in textureless or repetitive regions, and Eq. 6's best-of-four selection further biases it toward low uncertainty, so the central claim holds only conditionally.","rationale":"The reader's weakest_assumption identifies the same core issue: patch-based photometric consistency is treated as a proxy for geometric error, but this proxy is unreliable in textureless and repetitive-texture regions. The paper's own conclusion explicitly concedes degraded performance on large textureless areas, which is an in-scope admission that should weigh on the verdict. I further sharpen the concern by pointing to Eq. 6's selection of the four most consistent view pairs, which actively discards inconsistent observations and can produce low pseudo-labels even for geometrically erroneous surface points, as long as any subset of views happens to match. This makes the failure mode more systematic than simply 'textureless regions are hard.' However, this does not invalidate the method entirely: on textured scenes the results appear plausible, and the proposed concrete test would determine whether the limitation is confined to the admitted extreme cases or actually affects the core claim more broadly. Since the reader's verdict is already CONDITIONAL, and this concern reinforces that conditionality without demanding rejection, I leave the verdict unchanged.","tokens_in":13336,"tokens_out":5576,"duration_ms":53076,"concrete_test":"Construct a synthetic scene with a known GT mesh containing a large textureless planar region and a region with repeated texture. Train GURecon on sparse views, then compare the predicted uncertainty field (and the intermediate consistency G from Eq. 5-6) against the true distance of each surface point to the GT mesh, stratified by local patch variance. Compute the Spearman correlation between G and true error separately for low-variance and high-variance strata. If the correlation is near zero or negative in the low-variance or repetitive strata, the pseudo-label is invalid in those regimes. Additionally, report AUSE and AUSE3D computed only on test pixels whose local patch variance is below a threshold; if performance degrades to chance level there, the central claim must be conditioned on texture content.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the patch-based multi-view consistency score G computed in Eq. 5-6 is a valid, monotone proxy for true geometric error of the reconstructed surface. The uncertainty field is trained by L_distill (Eq. 7) to regress exactly this score. The paper's own conclusion admits degraded performance in large textureless areas (e.g., white walls), which is direct evidence that the proxy fails precisely where geometric uncertainty is most needed. Moreover, Eq. 6 aggregates only the four most consistent view pairs out of all available source views, explicitly discarding evidence of inconsistency. In repetitive-texture or low-gradient regions, some view pair will often match accidentally, so G can be low even when the reconstructed surface is substantially wrong. The uncertainty field then becomes confidently wrong, and the reported AUSE/AUSE3D gains may be driven primarily by textured, well-conditioned regions. This is a load-bearing concern because it targets the method's fundamental supervision signal, not just the evaluation protocol. It is acknowledged in the conclusion but never quantified or addressed in the experiments, so the generality of the central claim is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GURecon, a framework that learns a continuous 3D geometric uncertainty field for neural implicit surface reconstruction. The method locates zero-crossing surface points of the current SDF, computes patch-based multi-view photometric consistency scores as pseudo-labels, and distills these scores into an uncertainty field via an online distillation loss. A decoupled appearance branch is trained to remove view-dependent factors, and the uncertainty field is fine-tuned on the decoupled images. Experiments on DTU, BlendedMVS, and TNT report improved AUSE and AUSE3D scores relative to adapted NeRF uncertainty baselines, and the uncertainty is used for next-best-view selection in incremental reconstruction, where it improves Chamfer distance and PSNR. The authors claim this is the first geometry-focused, view-independent 3D uncertainty field for neural surfaces, learned without real geometric supervision.","tokens_in":13549,"tokens_out":4280,"duration_ms":39506,"significance":"If the central premise holds, the paper makes a useful contribution: it shifts uncertainty estimation for neural surface reconstruction from rendering-based pixel-wise scores to a 3D field tied to surface geometry, which is more natural for downstream tasks such as next-best-view planning and quality assessment. The plug-and-play extension to 2DGS and the consistent quantitative gains across three datasets are notable strengths, as is the promised code release. The key risk is that the pseudo-label itself is derived from the geometry being learned, so the uncertainty field can be confidently wrong when the surface is consistently wrong but photometrically consistent. The paper's own conclusion concedes degradation on large textureless regions, yet no experiment quantifies this failure mode. These issues are load-bearing for the claim of accurate geometric uncertainty estimation and require additional validation before the result can be fully trusted.","major_comments":[{"comment":"The load-bearing assumption is that the patch-based multi-view consistency score G^k is a monotone proxy for the geometric error of the reconstructed surface. Because this score is computed on the currently learned SDF and the uncertainty field is directly trained by L_distill to regress it, a consistently wrong geometry with photometrically consistent projections, such as a shifted low-texture plane or a repetitive-texture region, will receive low uncertainty. Eq. (6) aggregates only the four most consistent pairs, explicitly discarding evidence of inconsistency. The conclusion acknowledges degraded performance on large textureless areas, but the experiments never quantify how often the pseudo-labels are confidently wrong or how the reported AUSE/AUSE3D gains are distributed across textured and textureless regions. I request a direct validation of the pseudo-label: per-point correlation between G^k and the true distance-to-GT error, AUSE/AUSE3D stratified by texture gradient or by consistency-confidence bins, or a synthetic experiment with a known surface shift.","section":"3.2-3.3, Eqs. (5)-(7)"},{"comment":"The baselines are not compared in their native form: ActiveNeRF*, CFNeRF*, and Lee et al.* are structurally modified for SDF backends, while Bayes' Rays and Lee et al. are evaluated as post-hoc frameworks on the model trained by GURecon. As described, this makes the comparison depend on undocumented implementation choices (deferred to the supplementary material) and on the host model's geometry. The claim of superior performance relative to existing methods requires either evaluating all methods on the same SDF geometry with each method's own uncertainty computation, or additionally reporting the uncertainty quality of the original NeRF-based methods on their own reconstructions. Without this, the AUSE improvements cannot be attributed solely to the uncertainty model rather than to the integration strategy or the underlying reconstruction.","section":"4.1, Table 2"},{"comment":"AUSE measures only whether the ranking of predicted uncertainty matches the ranking of error; it does not assess calibration of the predicted uncertainty magnitudes. The qualitative statements in Figs. 3 and 6 that the learned uncertainties align well with GT error are not backed by a quantitative correlation or reliability analysis. Please report a rank correlation (e.g., Spearman) between predicted uncertainty and GT error at the surface-point level, and clarify how AUSE3D is computed, in particular which surface points or rays are used and how the continuous uncertainty field is evaluated on test views. This is needed to support the claim that GURecon models calibrated 3D geometric uncertainty rather than a relative confidence score.","section":"4.1, AUSE metrics"}],"minor_comments":[{"comment":"The row 'with smaller patch size 70' appears to be a typo; it should read '7' or '7 x 7' to be consistent with the reported patch size of 11 x 11.","section":"Table 3"},{"comment":"The reference 'Stereopsis, R. M. 2010' is incomplete and appears to be a garbled citation for Furukawa and Ponce's multi-view stereo paper; the entry should be corrected and expanded.","section":"References"},{"comment":"The symbol w_r' is used before its definition as the reflection of the view direction around the normal; please define it explicitly at first use and make the prime notation consistent.","section":"Eq. (8)"},{"comment":"The text 'Nf titerations' should read 'N_ft iterations'; the formatting makes the hyperparameter name unclear.","section":"3.5, Implementation Details"},{"comment":"The caption states that Bayes' Rays and Lee et al. share the same GT curve as ours, but the legend shows only one GT-error curve; please clarify the legend. Also, 'Tanks and Template' should be 'Tanks and Temples', and 'Soilder' in Table 2 should be 'Soldier'.","section":"Figure 5 / Section 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a computer vision journal and the empirical contribution is solid, but the central pseudo-label validity issue and the baseline adaptation protocol should be addressed before publication. No concerns about citation patterns or novelty disclosure beyond the need for a direct comparison with native baselines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is genuinely new: GURecon distills a continuous 3D geometric uncertainty field for an implicit SDF surface from multi-view patch-based photometric consistency, rather than the pixel-wise rendering-based uncertainty of Bayes' Rays or ActiveNeRF. Second, it works: the AUSE and AUSE3D numbers on DTU, BlendedMVS, and TNT are consistently better than the adapted baselines, and the extension to 2DGS suggests the distillation scheme is portable. The decoupling of view-dependent appearance to remove lighting interference is a sensible and effective addition. Credit where due: the authors know their limits—they state in the conclusion that large textureless regions degrade performance.\n\nThe soft spot is the pseudo-label itself. The consistency score is computed from the current geometry field, so it is self-referential. The stress-test note about Eq. 6 is fair: keeping only the four most consistent patch pairs discards evidence of inconsistency, biasing uncertainty low in repetitive or low-gradient regions. The paper acknowledges this in the conclusion but does not quantify it or offer a mitigation. That is a moderate weakness, not fatal—the method works on a range of realistic scenes, and the blind spot is disclosed.\n\nSmaller concerns: baselines are adapted with modifications described only in the supplement, and no code or variance estimates are provided. I cannot verify the project website's code from the manuscript. These are minor but worth pressing on.\n\nThe central claim—that you can learn a view-independent 3D geometric uncertainty field from consistency cues alone—holds up conditionally. It is a real within-subfield contribution and a serious referee should engage with it. My recommendation: accept for review, and ask the authors to validate the proxy, release code, and address the textureless-region limitation explicitly.","headline":"A solid, novel method for 3D geometric uncertainty in neural surface reconstruction, with a load-bearing but acknowledged weakness in textureless scenes.","tokens_in":14110,"tokens_out":3361,"would_cite":true,"duration_ms":25867,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"GURecon learns a per-point 3D geometric uncertainty field purely from multi-view photometric consistency, with no ground-truth geometry, and uses it to pick better next views.","keywords":["geometric uncertainty","neural surface reconstruction","multi-view photometric consistency","uncertainty distillation","implicit signed distance fields","next-best-view selection","view-dependent decoupling","3D reconstruction"],"falsifier":"Reconstruct a scene containing a large uniformly textured wall with GURecon, then displace the extracted surface a few centimeters parallel to the wall; if the uncertainty field assigns low uncertainty to the displaced surface because the photo patches still match, while the distance to the true surface is large, the photometric-consistency proxy is falsified.","tokens_in":13128,"feed_emoji":"🎯","tokens_out":9844,"duration_ms":79068,"temperature":0.7,"pith_summary":"GURecon sets out to show that a neural surface reconstruction's geometric quality can be scored per point without any ground-truth geometry, by learning a continuous 3D uncertainty field from multi-view photometric consistency. The method locates surface intersection points along rays in an implicit signed-distance field, measures how well small patches around those points agree across views, and trains a position-only uncertainty network to imitate those agreement scores. A decoupled appearance field removes view-dependent lighting and reflections before a final fine-tuning pass, so the uncertainty reflects geometry rather than illumination. If the paper is right, a reconstructed mesh carries its own reliability map, and that map can drive next-best-view selection while remaining independent of the viewing direction.","feed_headline":"Photo consistency becomes a 3D error map for neural surfaces","feed_subtitle":"GURecon distills multi-view patch agreement into a 3D uncertainty field that picks better next views.","key_machinery":"The machinery is an online distillation loop from geometry to uncertainty. At each training step, linear interpolation finds the ray segment where the SDF changes sign, giving surface points $T$; homography warping and SSIM over $11\\times11$ gray patches produce a consistency score $G^k$ (Eq. 5--6); and the uncertainty MLP $f_u(x)$, which takes only position, is trained with $L_{\\mathrm{distill}} = \\frac{1}{|R'|}\\sum_{r\\in R'} |f_u(p'_r) - G_r|$ (Eq. 7). A second decoupled branch writes color as $C = C_{vi} + C_{vd}$, and the rendered view-dependent part is subtracted from input images before the consistency scores are recomputed for a short fine-tuning stage, removing lighting and specular interference from the pseudo-labels.","core_discovery":"On the paper's own terms, the central discovery is that patch-based multi-view photometric consistency at surface points is a usable pseudo-label for geometric reconstruction error, and that a neural field can learn to predict that error in 3D. GURecon finds the zero-crossing of the SDF along each ray, approximates a local tangent plane at the intersection, warps image patches between views with a homography, scores agreement with SSIM on gray patches, keeps the four best-scoring pairs, and distills the resulting score into an uncertainty field $f_u(x)$ that depends only on position. Because the field is view-independent, the same surface point receives the same uncertainty from any direction, which distinguishes it from pixel-wise rendering-based uncertainty. The paper reports that the learned uncertainties align with true distance-to-ground-truth error and improve incremental reconstruction.","pith_inferences":["This suggests that the reliability of the pseudo-labels could be estimated online by the spread among the four selected patch pairs, letting the distillation loss down-weight internally inconsistent matches instead of averaging them.","A natural next step is to apply the same consistency-plus-decoupling recipe to pixel-wise rendering uncertainty, which could reduce lighting-induced overconfidence in NeRF-based uncertainty estimators.","In active scanning, the continuous field could allow direct gradient-based optimization of the next camera pose, rather than scoring a finite set of candidate views, since uncertainty is defined at every surface point.","The method's confidence in a region is bounded by the quality of the current reconstruction, so an incremental loop that re-distills after every added view would need recalibration to avoid compounding early overconfidence."],"forward_implications":["A scan can be annotated with per-point geometric confidence after training, letting a user see exactly which regions of a reconstructed mesh are trustworthy without comparing to any ground truth.","Next-best-view planning gains a view-independent signal: from any candidate camera, the uncertainty field can be rendered by sphere tracing, so the next viewpoint is chosen by summing predicted uncertainty over the visible surface.","The distillation schema transfers across surface representations, so new implicit or splatting-based reconstructors can inherit uncertainty estimates without redesigning their architecture.","Separating view-dependent appearance from geometry makes uncertainty scores robust to specular highlights and lighting changes across views, a failure mode of photometric-error-based methods.","On standardized benchmarks, the reported AUSE3D metric connects 2D image-space error curves to 3D geometric error, giving a direct way to compare uncertainty methods for reconstruction rather than novel-view synthesis."],"supporting_citations":[{"why":"Supplies the NeuS SDF and radiance field formulation, including volume rendering and the zero-level set, that GURecon builds on.","marker":"Wang et al. 2021"},{"why":"Establishes the photometric-consistency heuristic that patch agreement across views indicates geometric reliability.","marker":"Stereopsis 2010; Schönberger et al. 2016"},{"why":"Provides the homography-warping and patch-consistency machinery used to compute pseudo-labels at surface points.","marker":"Fu et al. 2022; Darmon et al. 2022"},{"why":"Supplies the view-dependent radiance decomposition that justifies the decoupled fields used to remove illumination before fine-tuning.","marker":"Verbin et al. 2022; Fan et al. 2023; Tang et al. 2023"},{"why":"Bayes' Rays is the rendering-based uncertainty baseline against which 3D geometric uncertainty is compared.","marker":"Goli et al. 2023"},{"why":"ActiveNeRF is the uncertainty-guided next-best-view baseline that GURecon must beat in incremental reconstruction.","marker":"Pan et al. 2022"}],"fun_headline_variants":["Patch agreement teaches neural surfaces their own error","Neural surfaces learn 3D uncertainty from photo consistency","No ground truth needed: GURecon predicts geometry error in 3D","Multi-view patch scores become a 3D geometric uncertainty field","Self-distilled 3D uncertainty for reconstructions from photos"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method treats photometric consistency — the same surface patch looking alike across several views — as a sign that the surface is geometrically correct, which fails when a wrong surface sits on blank or repeating texture and still looks consistent.","fun_headline_variants_meta":{"raw":{"variants":["Patch agreement teaches neural surfaces their own error","Neural surfaces learn 3D uncertainty from photo consistency","No ground truth needed: GURecon predicts geometry error in 3D","Multi-view patch scores become a 3D geometric uncertainty field","Self-distilled 3D uncertainty for reconstructions from photos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1376,"prompt_tokens":923,"completion_tokens":453,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":369}},"tokens_in":539,"tokens_out":453,"duration_ms":4180,"temperature":1.0,"reasoning_tokens":369,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:46:49.684660+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reconstruct a scene containing a large uniformly textured wall with GURecon, then displace the extracted surface a few centimeters parallel to the wall; if the uncertainty field assigns low uncertainty to the displaced surface because the photo patches still match, while the distance to the true surface is large, the photometric-consistency proxy is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the NeuS SDF and radiance field formulation, including volume rendering and the zero-level set, that GURecon builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the photometric-consistency heuristic that patch agreement across views indicates geometric reliability."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the homography-warping and patch-consistency machinery used to compute pseudo-labels at surface points."},{"cited_title":"T.; and Srinivasan, P","cited_arxiv_id":null,"evidence_quote":"Supplies the view-dependent radiance decomposition that justifies the decoupled fields used to remove illumination before fine-tuning."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ActiveNeRF is the uncertainty-guided next-best-view baseline that GURecon must beat in incremental reconstruction."}],"review_version":1}