{"id":"3d0ad1b5-3212-4ab2-933d-1e23956b63b8","arxiv_id":"2412.15909","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"A curvature-based supervision signal for neural distance fields is proposed and tested on LiDAR mapping and Monte Carlo localization.","lead":"This paper proposes CCNDF, a method that trains neural distance fields from LiDAR scans by using the curvature of the field's contour lines to estimate signed distances. A reader might care because the method aims to make large-scale mapping and localization from self-supervised neural fields more accurate.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The derivation of Eq. (8) assumes the LiDAR endpoint e_i lies on the osculating circle centered at the curvature center of the nearest surface point F; for general surfaces this is false, so the proposed distance is biased and the central claim is unsupported.","rationale":"The reader identifies the crucial flaw: Eq. (7) replaces ||e_i - c_l|| with r_l, the radius of curvature at the nearest point F. This is only valid when e_i lies on the same osculating circle as F. In general, the NDF isoline through x_l has curvature center c_l, which is the curvature center of the surface at F, but the ray endpoint e_i is an arbitrary point on the surface. For non-circular shapes, ||e_i - c_l|| does not equal r_l, so the computed r_l and hence d_hat_l are wrong. This is not a boundary case; it is the generic situation. The paper's own limitation section admits the assumption and appeals to sampling density, but a near-surface analysis shows d_hat_l reduces to d_l cos(theta), the LocNDF first-order projection, so the claimed improvement over LocNDF vanishes exactly where the loss weights concentrate. The empirical comparisons (Tables 1–2) are not accompanied by code or error bars, and Table 1 has implausible RMSE/MAE relationships (e.g., ray distance RMSE 4.42 vs MAE 1.31), so they cannot rescue the theory. The core geometric innovation is therefore unsupported, and the reader's REJECT verdict is appropriate.","tokens_in":12513,"tokens_out":7248,"duration_ms":56342,"concrete_test":"Use a 2D ellipse (e.g., x^2/4 + y^2 = 1) as a ground-truth surface with a known SDF. Pick a query point x_l on an offset curve, compute its nearest point F, the curvature center c_l of the offset, and the true radius r_l = ||F - c_l||. Choose a ray whose endpoint e_i is a different point on the ellipse (not F) and compute ||e_i - c_l||. If ||e_i - c_l|| differs from r_l, then Eq. (7)'s substitution of r_l^2 = ||e_i - c_l||^2 is wrong. Now compute d_hat_l from Eq. (8) and compare with the true signed distance ||x_l - F||. Repeat for rays hitting the ellipse at various angles and for sample points approaching the surface. If d_hat_l fails to converge to the true distance (except when x_l is exactly at e_i), then the supervision signal is biased and the claimed geometric advantage is not real.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption in Eq. (7)–(8) is that the LiDAR endpoint e_i lies on the osculating circle centered at c_l with radius r_l, where c_l is the curvature center of the NDF isoline at the query point x_l. For an SDF, c_l equals the curvature center of the surface at the nearest point F to x_l, but e_i is the point where the ray happens to hit the surface. The distance ||e_i - c_l|| is not equal to the radius of curvature r_l at F unless the surface is a circle or the ray passes through the curvature center. The cosine rule therefore computes the wrong r_l, and d_hat_l = R_l - r_l is not the signed distance to the nearest surface. This is not a minor edge case: for any surface with varying curvature, the ray endpoint and the foot point are different, and the error persists. The paper's Section 5 'Geometric Limitation' acknowledges the assumption, but claims it is mitigated by sampling near the surface. However, as x_l approaches the surface, d_hat_l tends to d_l cos(theta), the LocNDF first-order estimate, so the proposed method does not reduce to the true distance and the claimed advantage over LocNDF disappears precisely in the near-surface region that is weighted most heavily. Thus Eq. (8) does not provide a more accurate supervision signal, and the central claim is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CCNDF, a self-supervised method for training neural distance fields (NDF) from LiDAR sequences. The central idea is to compute, for each sampled ray point, a 'curvature-constrained' signed distance estimate that uses the second-order differential structure of the current NDF, and to supervise the network with this estimate. The authors claim that this yields more accurate mapping and localization than ray distance and LocNDF's first-order estimate. The evaluation consists of qualitative mapping results on Apollo and KITTI datasets and quantitative 2D Monte Carlo localization results on five sequences, compared with LocNDF and several classic baselines.","tokens_in":12839,"tokens_out":9461,"duration_ms":79350,"significance":"If the proposed geometric estimator were correct, using curvature information to supervise neural distance fields would be a meaningful extension of first-order methods, and the localization benchmark could provide practical value. The paper is clearly written in parts and the authors identify a genuine limitation of existing supervision signals. However, the derivation of the estimator is not valid for general surfaces, and the paper's own limitation section does not resolve the issue. The idea is worth exploring, but the current formulation and evaluation do not support the central claim.","major_comments":[{"comment":"The derivation sets ||e_i - c_l|| = r_l, where e_i is the LiDAR ray endpoint, c_l is the curvature center of the NDF isoline at the query point x_l, and r_l is the radius of curvature at the nearest surface point F. This equality holds only if the ray endpoint lies on the same circle as F, i.e., only for circular (or spherical) surfaces. For a general surface, e_i is not equal to F, and the distance from e_i to c_l is not r_l; therefore the cosine rule in Eq. (7) does not produce the radius of curvature at the nearest surface point, and Eq. (8) is not an approximation of the true signed distance to the nearest surface. The toy example in Fig. 2(b) is a circle, where the assumption happens to hold, so it cannot validate the general claim. The 'Geometric Limitation' paragraph in Section 5 acknowledges that the ray may miss the curvature circle at F but asserts that near-surface sampling mitigates the error; however, expanding Eq. (8) for small ray distance d_l gives d_hat_l ≈ d_l cosθ = n_l^T(e_i - x_l), which is exactly LocNDF's first-order estimate. Thus, in the near-surface regime where sampling weight is highest, the proposed method reduces to the method it claims to improve, and the claimed advantage is not supported.","section":"Section 3.1, Eqs. (6)-(8) and Fig. 2(a)"},{"comment":"The formulas for mean curvature and for the 'radius of curvature' of the NDF are not correct as stated. The standard mean curvature of an implicit surface is κ = (∇f^T H(f) ∇f - ||∇f||^2 trace(H)) / (2||∇f||^3), but Eq. (4) contains an extra factor of 2 in the '2trace(H)' term. Equation (5) writes the divergence of the normalized gradient as a norm of a 'gradient' with unclear notation; the divergence is a scalar, and the formula as printed is not the divergence. Moreover, for a general surface the mean curvature is not the reciprocal of a single radius of curvature; the 'radius of curvature of an isoline' is only well defined for circles or spheres. Since R_l is a load-bearing quantity in Eqs. (7)-(8), the geometric interpretation of the method is not established for surfaces with anisotropic or saddle curvature.","section":"Section 3.1, Eqs. (4)-(5)"},{"comment":"The evaluation protocol is described only at a high level, and the reported numbers need more support. The paper states that metrics are averaged over 5 runs, but no variances are given; Table 2 shows identical RMSE values for the proposed method at 5 cm and 10 cm thresholds for Seq-1 and Seq-2 (1.4 and 1.9), which is unusual and not explained. In addition, no ablation study isolates the contribution of the curvature term from the regularizers λ1, λ2, λ3, or from the weighting scheme, so it is not clear that the empirical improvement, if any, is due to the proposed distance estimate. Since the paper's central claim is state-of-the-art performance, this lack of support is problematic.","section":"Section 4, Tables 1 and 2"}],"minor_comments":[{"comment":"The positional encoding definition does not specify the frequencies ω_i; please define them explicitly.","section":"Eq. (1)"},{"comment":"The log-linear sampling formula is not typeset clearly; the intended expression for t_l is difficult to parse and should be written with unambiguous parentheses and exponents.","section":"Eq. (2)"},{"comment":"The text alternates between 'isolines' and 'surfaces' without clarifying that the NDF is a 3D scalar field and that the relevant objects are level surfaces, not 2D curves.","section":"Section 3.1"},{"comment":"The acronym 'ROC' is used extensively but is only spelled out in the caption; please define it at first use in the text.","section":"Fig. 2"},{"comment":"Reporting mean plus/minus standard deviation and the number of runs that converged would improve the interpretability of the localization results.","section":"Tables 1 and 2"},{"comment":"The claim 'state-of-the-art' should be tempered because the comparison set is limited to LocNDF and classical baselines; no comparison with other recent neural mapping approaches is provided.","section":"Section 4"}],"recommendation":"reject","confidential_remarks":"The paper's central contribution rests on a geometric derivation that is invalid for general surfaces; this is not a matter of presentation but of the core estimator. The self-supervised nature and the lack of ablations further weaken the empirical claims. While the idea of using curvature for NDF supervision is interesting, the current formulation cannot be fixed with local edits and the manuscript would need a re-derivation of the distance estimate and a much more thorough evaluation. I recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The paper introduces something genuinely new: supervising a neural distance field from LiDAR using second-order curvature of the NDF, not just ray distance or first-order gradient. That is a real gap in the cited LocNDF/SHINE/i-SDF line. The paper also does the honest thing by stating its geometric limitation in Section 5. But the central formula, Eq. (8), is not a valid distance to the nearest surface for general geometry, and the evaluation cannot compensate.\n\nThe derivation starts from R_l, the radius of curvature of the NDF isoline at the query point. For a true distance field, the curvature center of that isoline is also the curvature center of the surface at the nearest foot point F. The paper then wants r_l, the surface radius at F. The cosine rule in Eq. (7) instead sets r_l = ||e_i - c_l||, where e_i is wherever the LiDAR ray happens to hit the surface. Those are equal only when e_i is on the osculating circle of F—i.e., for circles/spheres or rays through the curvature center. General LiDAR rays hit surfaces at other points, so ||e_i - c_l|| is not r_l. The paper's Section 5 acknowledges the condition, but the mitigation is hand-wavy: as the query point approaches the surface, d_hat from Eq. (8) tends to d_l cos(theta), which is exactly LocNDF's estimate. So the claimed advantage vanishes precisely in the near-surface region that their weighting emphasizes. That is a load-bearing flaw, not a minor edge case.\n\nI can't verify the empirical claims. There is no code, no error bars, and mapping is evaluated only qualitatively. The MCL numbers are intriguing, but the comparison lacks the statistical detail needed to trust improvements of a few centimeters. Eq. (5) also looks dimensionally odd as written (the norm of the gradient of the normalized gradient is not the mean curvature; there are missing factors), and the notation is loose throughout.\n\nWhat's genuinely worth keeping is the idea. If the authors can correct the geometry—or reframe the contribution as a curvature-aware heuristic with a proper derivation for smooth surfaces—it could be a useful paper. As it stands, the central claim is unsupported. I'd send it out anyway, because the novelty and the application area deserve referee time, but it needs major revision or a fundamentally different derivation before I'd trust the results.","headline":"The curvature-supervision idea is new, but Eq. (8) does not compute distance to the nearest surface for general geometry, and the reported gains are not verifiable.","tokens_in":13349,"tokens_out":4198,"would_cite":false,"duration_ms":39261,"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":"Supervising a neural distance field with a curvature-derived signed-distance estimate improves LiDAR mapping and localization over ray-distance and first-order normal baselines.","keywords":["neural distance fields","curvature estimation","signed distance supervision","LiDAR mapping","self-supervised learning","implicit neural representations","Monte Carlo localization"],"falsifier":"On a synthetic scene with known geometry (a sphere of radius 1 and a flat plane), simulate LiDAR rays that hit the surface at points different from the nearest point to samples along the ray; compute the proposed estimate $\\hat{d}_\\ell$ at each sample and compare with the analytic signed distance. If the estimate's error on off-center sphere rays or on flat regions is not smaller than the errors of raw ray distance and LocNDF's normal projection, the central geometric advantage asserted by the paper would be refuted.","tokens_in":12284,"feed_emoji":"🗺️","tokens_out":8645,"duration_ms":64824,"temperature":0.7,"pith_summary":"Neural distance fields (NDFs) — neural networks that map a 3D point to its signed distance from the nearest surface — are attractive for LiDAR mapping but normally need ground-truth distance values that do not exist for outdoor scenes. The paper's claim is that a better self-supervised training signal can be derived from the field's own curvature: at any query point along a LiDAR ray, the signed distance to the surface is approximated as the difference between the radius of curvature of the NDF isoline at that point and the radius of the osculating circle through the ray's surface endpoint. The proposed estimate, $\\hat{d}_\\ell = R_\\ell - \\sqrt{d_\\ell^2 + R_\\ell^2 - 2 R_\\ell n_\\ell^\\top (e_i - x_\\ell)}$, is shown on a toy example to land at 0.98 where the true signed distance is 1, while ray distance gives 1.41 and LocNDF's normal-projected distance gives 1.75. If the claim holds, the same raw LiDAR scan yields sharper maps (cars, trees, building edges) and more reliable Monte Carlo localization without any new sensors or ground-truth meshes. The paper's experiments support this with qualitative reconstructions on KITTI and Apollo ColumbiaPark-3 and consistently lower RMSE/MAE localization numbers than the compared baselines.","feed_headline":"Curvature-corrected distance sharpens LiDAR neural maps","feed_subtitle":"Replacing ray distance with isoline curvature gives finer maps and lower localization error.","key_machinery":"The load-bearing object is the curvature-constrained distance estimate of Equation (8), $\\hat{d}_\\ell = R_\\ell - \\sqrt{d_\\ell^2 + R_\\ell^2 - 2 R_\\ell n_\\ell^\\top (e_i - x_\\ell)}$. The radius of curvature $R_\\ell$ of the NDF isoline is computed from the divergence of the normalized gradient of the field, $R_\\ell = \\left\\|\\nabla \\cdot \\left(\\frac{\\nabla D}{\\|\\nabla D\\|}\\right)\\right\\|^{-1}$; the cosine rule in the triangle with the ray endpoint $e_i$ and the curvature center $c_\\ell$ then yields the radius $r_\\ell$ of the circle through the endpoint. Its role is to turn the nearest-surface direction $n_\\ell$ (a first-order quantity) into a scalar distance that respects the curvature of the surface's level sets, and that scalar supervises the MLP through the weighted loss over log-linearly sampled ray points.","core_discovery":"The central discovery is that the signed distance from a query point to the nearest surface can be read off the curvature structure of the learned distance field itself. Because the field's level sets are claimed to be concentric with the surface, the true distance is the difference between the radius of curvature $R_\\ell$ of the isoline through the query point and the radius $r_\\ell$ of the circle through the ray endpoint on the surface, both referred to the same curvature center; the paper derives $r_\\ell$ from the cosine rule in the triangle formed by the query point, the ray endpoint, and that center (Eq. 7), giving $\\hat{d}_\\ell = R_\\ell - r_\\ell$ (Eq. 8). The paper argues that this second-order supervision is more faithful than raw ray distance, which ignores the geometry of the rest of the surface, and than LocNDF's first-order normal projection, which assumes locally linear surfaces and overshoots on curved ones. It then trains the same SIREN-based MLP with the new distance as the supervision target and reports that the resulting maps capture finer detail and that 2D Monte Carlo localization on five sequences achieves lower RMSE/MAE than the LocNDF, IR-MCL, SRRG, and AMCL baselines.","pith_inferences":["If the curvature principle carries over to signed distance functions of general surfaces, the same Equation (8) could in principle be applied to any range sensor — RGB-D cameras, sonar — where ray endpoints and field curvature are available, not just LiDAR point clouds.","The geometric derivation suggests a testable scaling law: the error of ray-distance supervision should grow with local curvature, while the proposed estimate should stay bounded, so sorting scenes by curvature could predict which scenes benefit most.","A natural extension the authors do not develop is online incremental training: since the curvature estimate depends on the currently learned field, streamed LiDAR would need a stabilization scheme, though the log-linear ray sampling they use may already provide one.","The concentric-isolines argument is exact for spheres and cylinders; for general surfaces it is an approximation whose quality likely degrades with the ratio of local curvature radius to the distance between the query point and the ray endpoint, a quantity that could be computed from the learned field itself."],"forward_implications":["Mapping pipelines that currently supervise neural fields with raw ray distance could collect the same LiDAR scan and reconstruct finer geometry — cars, trees, building edges — without any new sensors or ground-truth meshes.","Because the estimate is expressed through the field's own curvature, the supervision becomes a self-consistent second-order constraint: better fields give better distances, which feed better fields, instead of relying on a fixed first-order projection.","The improved map representation transfers to downstream tasks: 2D Monte-Carlo localization reports lower RMSE and MAE on five sequences, and the error values saturate rather than growing with the threshold from 5 to 20 cm.","The approach extends self-supervised NDF supervision from first-order properties (direction to nearest surface) to second-order properties (curvature), opening the same machinery to other range sensors such as depth cameras."],"supporting_citations":[{"why":"supplies the backbone MLP, log-linear ray sampling, nearest-surface normal estimate, and the localization protocol that the proposed supervision must beat.","marker":"[30]"},{"why":"provides the implicit-curve curvature formula used to compute the radius of curvature R_l of the NDF isoline.","marker":"[8]"},{"why":"defines the SHINE-based MLP and TSDF structure for large-scale LiDAR mapping, and the ray-distance supervision baseline.","marker":"[38]"},{"why":"provides the first-order normal-distance supervision method whose overestimation on curved surfaces motivates the curvature correction.","marker":"[19]"},{"why":"supplies the TSDF integration and the near-surface weighting w_l used to stabilize training.","marker":"[27]"},{"why":"provides the SIREN periodic-activation architecture that represents the neural distance field.","marker":"[23]"},{"why":"defines the localization benchmark sequences and the IR-MCL baseline against which localization results are compared.","marker":"[11]"},{"why":"provides the KITTI dataset used for mapping and localization evaluation.","marker":"[7]"},{"why":"provides the Apollo Southbay ColumbiaPark-3 dataset used for large-scale mapping evaluation.","marker":"[13]"}],"fun_headline_variants":["Curvature-aware distance fields boost LiDAR mapping and localization","Second-order SDF supervision sharpens neural maps from LiDAR","Using isoline curvature refines LiDAR neural distance fields","Curvature-constrained NDF cuts localization error on LiDAR maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The derivation assumes the LiDAR ray endpoint lies on the same curvature circle as the nearest surface point to the query point, which is exact for spheres and other circular geometries but not for general surfaces.","fun_headline_variants_meta":{"raw":{"variants":["Curvature-aware distance fields boost LiDAR mapping and localization","Second-order SDF supervision sharpens neural maps from LiDAR","Using isoline curvature refines LiDAR neural distance fields","Curvature-constrained NDF cuts localization error on LiDAR maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000582,"raw_usage":{"total_tokens":2766,"prompt_tokens":996,"completion_tokens":1770,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":1698}},"tokens_in":612,"tokens_out":1770,"duration_ms":10398,"temperature":1.0,"reasoning_tokens":1698,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:58:31.323237+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic scene with known geometry (a sphere of radius 1 and a flat plane), simulate LiDAR rays that hit the surface at points different from the nearest point to samples along the ray; compute the proposed estimate $\\hat{d}_\\ell$ at each sample and compare with the analytic signed distance. If the estimate's error on off-center sphere rays or on flat regions is not smaller than the errors of raw ray distance and LocNDF's normal projection, the central geometric advantage asserted by the paper would be refuted.","supporting_citations":[{"cited_title":"IEEE Robotics and Automation Letters (2023) 1, 2, 3, 4, 5, 8, 10, 11, 12, 13","cited_arxiv_id":null,"evidence_quote":"supplies the backbone MLP, log-linear ray sampling, nearest-surface normal estimate, and the localization protocol that the proposed supervision must beat."},{"cited_title":"Computer Aided Geometric Design 22(7), 632–658 (2005) 7","cited_arxiv_id":null,"evidence_quote":"provides the implicit-curve curvature formula used to compute the radius of curvature R_l of the NDF isoline."},{"cited_title":"In: 2023 IEEE Inter- national Conference on Robotics and Automation (ICRA)","cited_arxiv_id":null,"evidence_quote":"defines the SHINE-based MLP and TSDF structure for large-scale LiDAR mapping, and the ray-distance supervision baseline."},{"cited_title":"system30, 39 1, 2, 3, 4","cited_arxiv_id":null,"evidence_quote":"provides the first-order normal-distance supervision method whose overestimation on curved surfaces motivates the curvature correction."},{"cited_title":"Sensors22(3), 1296 (2022) 5, 8","cited_arxiv_id":null,"evidence_quote":"supplies the TSDF integration and the near-surface weighting w_l used to stabilize training."},{"cited_title":"Advances in neural information processing systems 33, 7462–7473 (2020) 1, 3, 4, 10","cited_arxiv_id":null,"evidence_quote":"provides the SIREN periodic-activation architecture that represents the neural distance field."},{"cited_title":"IEEE Robotics and Automation Letters 8(3), 1627–1634 (2023) 11, 12, 13","cited_arxiv_id":null,"evidence_quote":"defines the localization benchmark sequences and the IR-MCL baseline against which localization results are compared."},{"cited_title":"In: 2012 IEEE conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"provides the KITTI dataset used for mapping and localization evaluation."},{"cited_title":"In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition","cited_arxiv_id":null,"evidence_quote":"provides the Apollo Southbay ColumbiaPark-3 dataset used for large-scale mapping evaluation."}],"review_version":1}