{"id":"f8b27bb6-3931-474c-ba26-722a4c98c24f","arxiv_id":"2502.06047","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A vector-valued neural field that learns the exact shortest path to a surface reconstructs open 3D shapes from point clouds while recovering distance and gradient jointly.","lead":"This paper trains a neural network to predict, for any point in space, the vector that points to the nearest point on a target surface, together with the distance and gradient hidden in that vector. The authors prove a convergence property for this representation and test it on reconstruction of garments, cars, and indoor scenes from point clouds.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2 is proved for the idealized loss (23), not for the Ltotal actually minimized in training, and its Appendix A.2 proof lacks the Lipschitz regularity needed for the mean-value step; the advertised guarantee is therefore not established for the reported model.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: Theorem 2 is proved for the idealized undetached loss LNSP, while the reported training minimizes Ltotal with a minimal-area term and a detached outer gradient, and the proof of Theorem 2 contains an unjustified mean-value step applied to an H1 function. I read the paper in good faith: the NSP representation and the conditional H1-convergence result (Theorem 1) are meaningful contributions, and the experimental comparison, though lacking error bars and code, supports a conditional rather than a rejected verdict. The main concern is that the central theoretical guarantee is not directly applicable to the actual trained objective, and the proof of the idealized theorem is incomplete. An ablation with the exact LNSP loss would settle whether the practical claims rely on the untheorized components; if the exact-loss model performs comparably, the concern is mitigated. Because these are addressable gaps rather than demonstrated failures, the reader's CONDITIONAL verdict stands without change.","tokens_in":19256,"tokens_out":26135,"duration_ms":251227,"concrete_test":"Retrain the published architecture on the partial-cylinder and shirt datasets with the exact idealized loss LNSP (23), i.e. no minimal-area term and no gradient detachment, and compare the H1 error of the learned dθ against the analytic distance field and against the Ltotal-trained model. If the exact-loss model matches or improves on the Ltotal model, the extra terms are not essential and the theorem's scope is adequate; if it is materially worse or fails to converge, the reported reconstructions depend on components of Ltotal that Theorem 2 does not cover.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that a global minimizer of the proposed loss is the ESP. The theorem behind this (Theorem 2) is stated for LNSP (23) and requires all three terms to be exactly zero at the minimizer. In actual training, however, the objective is Ltotal (26), which appends a minimal-area term λMA∫δϵ(dθ) and implements the shortest-path term with a detached outer d̂θ (20). The area term changes the global-minimizer set: no result in the paper shows that minimizers of Ltotal are minimizers of LNSP, so the ESP characterization need not transfer. The detached gradient leaves the loss value unchanged but produces a surrogate gradient, so the optimization trajectory is not the gradient flow of LNSP either. Furthermore, the proof of Theorem 2 in Appendix A.2 applies the multivariate mean-value theorem to d*, which is only known to be in H1(Ω); in R3 this is insufficient for pointwise differentiability along the segment [x, x−d(x)∇d(x)], and no Lipschitz regularity (e.g. from |∇d*|≤1 a.e.) is established to justify the bound |d*(a)−d*(b)|≤|a−b|. Thus the proof is incomplete as written. These two gaps leave the advertised convergence/ESP guarantee without a direct link to the network whose reconstructions are reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Neural Shortest Path (NSP), a vector-valued implicit neural representation Fθ = dθ Gθ whose magnitude approximates the unsigned distance function to a surface and whose unit direction approximates its gradient. The method decomposes the network output via magnitude-direction decomposition (MDD) and uses a variable-splitting training loss (manifold, gradient matching, and shortest-path terms) plus a minimal-area regularization. The authors claim two theoretical results: (i) Theorem 1 says that under a.e. convergence of the network output and vanishing gradient-matching loss, the magnitude converges in H1 norm; (ii) Theorem 2 says that a global minimizer of the idealized loss LNSP is the exact shortest path (ESP) for some surface Γ' containing the target Γ. They also introduce a shortest-path-based surface extraction algorithm and report experiments on synthetic, garment, car, indoor-scene, noisy, and sparse point clouds, showing qualitative and quantitative improvements over NDF, CAP, and CSP.","tokens_in":19611,"tokens_out":7548,"duration_ms":75503,"significance":"If the theoretical claims were fully established, the paper would make a valuable contribution: a single-network representation that simultaneously recovers distance and gradient, with a convergence guarantee in H1 avoiding higher-order derivatives, and a surface extraction method that bypasses the kinks at the zero level set. The empirical evaluation is extensive and suggests that the proposed loss is practically effective for open and non-watertight surfaces. Credit should be given for the explicit ESP construction, the variable-splitting formulation, and the reproducibility-oriented experimental reporting (datasets, baselines, and implementation details are given). However, the theoretical results as stated are consistency statements derived from loss terms designed to encode exactly the ESP properties; they do not certify that the trained model converges to the target surface. Moreover, the main theorem's proof has a regularity gap, and the theorem applies to a loss that is not the one minimized in the reported experiments. These issues are fixable in principle, but they currently leave the advertised guarantees without a direct link to the deployed method.","major_comments":[{"comment":"Theorem 1 as stated is false: a sequence of nonzero a.e. H1 functions can converge a.e. to a limit that is zero on a set of positive measure (e.g., F_n = (1/n, 0, 0) on Ω), so the proof's assertion that 'the limit point F is also nonzero a.e.' is unjustified. The subsequent convergence proof for G_n = F_n/||F_n|| divides by d = ||F||, which is not well-defined on the zero set of the limit. The theorem needs an explicit assumption that the a.e.-limit F is nonzero a.e., or a separate handling of the zero set; as written, it does not establish the advertised H1 convergence.","section":"Theorem 1, Section 3.3 and Appendix A.1"},{"comment":"The proof of Theorem 2 applies the multivariate mean-value theorem to d*, which is only known to lie in H1(Ω). In three dimensions, H1 functions are not necessarily continuous or differentiable at every point on the segment [x, x − d(x)∇d(x)], and the inequality |d*(a) − d*(b)| ≤ |a − b| does not follow from the weak relation |∇d*| = 1 a.e. without Lipschitz regularity or a viscosity/regularity argument. Thus the central comparison d*(x) ≤ d(x) and its reverse are not established. The proof needs an additional regularity statement for minimizers, e.g., showing that LSP = 0 plus LGM = 0 implies d* is the viscosity solution of the eikonal equation, not merely an H1 function.","section":"Theorem 2, Appendix A.2"},{"comment":"The global-minimum statement is proved for LNSP in (23)/(24), but the reported models are trained with Ltotal (26), which adds the minimal-area term λMA ∫ δϵ(dθ) dx, and the shortest-path term is implemented with a detached outer gradient (20). No result shows that minimizers of Ltotal coincide with minimizers of LNSP or that the surrogate gradient in (20) preserves the global-minimum structure. Consequently, the theoretical guarantee does not directly apply to the objective actually minimized in the experiments, leaving the paper's central claim unsupported for the trained model.","section":"Section 3.3 vs. Section 3.2, Eq. (26) and (20)"},{"comment":"Theorem 2 does not prove that the recovered surface equals the target surface: it only shows Γ' (the zero set of the minimizer) contains Γ, so the minimizer could be the distance function of any superset surface. The minimal-area term is thus essential for fidelity, but it is not part of the theoretical analysis. In addition, the paper assumes without proof that an optimal solution exists in the admissible set {F ∈ H1(Ω, R3) : F ≠ 0 a.e.}; the functional has no coercivity in ∇F itself and the nonzero constraint is not closed, so the infimum may not be attained. The theorem is vacuous unless existence is established.","section":"Theorem 2 statement and existence"}],"minor_comments":[{"comment":"There is a typo in the phrase 'demonstrate the the H1 norm convergence'; please remove the duplicated article.","section":"Section 3.3"},{"comment":"The loss functional is introduced twice, once in Eq. (14) as the 'proposed loss functional' and again in Eq. (23) as LNSP with identical terms; please consolidate the notation to avoid confusion.","section":"Section 3.2"},{"comment":"The noise standard deviations reported for the shirt data (σ = 0.002, 0.004, 0.005) are very small relative to the normalized domain [−1,1]^3; a brief note on the scale of the point-cloud coordinates would help the reader interpret the robustness results.","section":"Section 4.4"},{"comment":"There is a typo in the word 'utlizing' and the softplus formula would be clearer with explicit parentheses: σ(x) = (1/β) log(1 + e^(βx)).","section":"Appendix B"},{"comment":"The notation 'a.e.' is used extensively (e.g., Eq. (24), Theorem 1) before it is first defined; please define it at first use in Section 2.1 or 3.1.","section":"Section 3.2 and 3.3"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a practically promising method with strong empirical results, but the theoretical sections need substantial revision. The gap between Theorem 2 and the training objective (Ltotal with detached gradients), combined with the regularity gap in Appendix A.2, means the central guarantee is not currently established for the reported system. These issues are likely fixable with additional assumptions or a scoped theorem, so I do not see grounds for rejection if the authors are willing to revise. It may be worth consulting a PDE or geometric measure theory expert for the regularity of distance functions to repair Appendix A.2."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is worth reading. The core idea—parameterize F = d∇d as a single vector field, decompose it into magnitude and direction, and train with a shortest-path loss—is a real step beyond CSP, NDF, and CAP. It directly targets the gradient distortion at kinks that plagues distance-field networks, and the H1 convergence theorem (Theorem 1) is a clean conditional result: if the gradient-matching residual vanishes and the network converges, the magnitude converges in H1. The shortest-path surface extraction is also sensible and avoids relying on noisy gradients near the zero set. Empirically, the method is competitive or better on the reported open-surface benchmarks, with visually smoother outputs and qualitative robustness to noise and sparsity.\n\nThe soft spots are real but addressable. Most importantly, Theorem 2 proves that a global minimizer of LNSP (eq. 23) is an ESP, but training minimizes Ltotal (eq. 26), which adds a minimal-area term and detaches the gradient in the shortest-path loss. No theorem connects minimizers of Ltotal to the ESP characterization, and the detached gradient means the optimization trajectory is not even the gradient flow of LNSP. The abstract's claim that \"its global minimum is the ESP\" is therefore stronger than what is shown for the actual method. Second, the proof of Theorem 2 in Appendix A.2 applies a mean-value step to d*, which is only known to be in H1; without Lipschitz regularity along the path, the inequality |d*(a)-d*(b)| ≤ |a-b| needs justification. This is likely fixable, but the proof is incomplete as written. There is also a small gap in the Theorem 1 proof: an a.e. limit of H1 functions that are nonzero a.e. need not be nonzero a.e. The eikonal constraint probably saves the argument, but the step is not supplied. Finally, the \"substantial improvements\" wording overstates the tables—on cars and indoor scenes NSP is roughly on par with CAP—and the robustness comparisons are qualitative only, with no error bars or code.\n\nOverall, the representation and the shortest-path loss are solid contributions, and the theory is close to being right but needs tightening. I would send this to peer review and ask for a revised version that either proves the global-minimum result for the loss actually minimized or clearly states the limitation, and that adds quantitative noise/sparsity metrics. I would cite it for the F = d∇d idea, not for its theorems until they are cleaned up.","headline":"The F = d∇d representation is a genuine advance for open-surface INR, but the ESP guarantee is proved for a loss that is not the one being trained.","tokens_in":20073,"tokens_out":6988,"would_cite":true,"duration_ms":71178,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D17","68T07","35F21"],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural field that learns the exact shortest path to a surface, not just its distance, can reconstruct open shapes from noisy point clouds.","keywords":["surface reconstruction","point clouds","implicit neural representation","unsigned distance function","exact shortest path","variable splitting","eikonal equation","H1 convergence"],"falsifier":"Train the full network with the practical loss on a clean hemisphere point cloud, then compare the learned field against the ground-truth distance $d$ and shortest path $d\\nabla d$ on a fine grid away from the surface. If, at convergence, the $H^1(\\Omega)$ error $\\|d_\\theta-d\\|_{H^1(\\Omega)}$ or the shortest-path error $\\|F_\\theta-d\\nabla d\\|_{L^2(\\Omega\\setminus\\Gamma)}$ is bounded away from zero even as the loss approaches zero, then the reported guarantee does not transfer to the deployed training objective.","tokens_in":19018,"feed_emoji":"🧭","tokens_out":10270,"duration_ms":92499,"temperature":0.7,"pith_summary":"The paper is trying to establish that a neural network can learn, directly from a point cloud, the exact shortest path to the unknown surface: the vector field that points each query toward its nearest surface point and whose length is the distance to the surface. Its central object is the identity $F(x)=d(x)\\nabla d(x)$, so learning $F$ means learning the distance and its gradient at once. The authors prove that when this vector field is decomposed into magnitude and unit direction and trained with their three-part loss, the magnitude converges in the $H^1(\\Omega)$ norm, meaning the gradient converges too, and the global minimizer of the idealized loss is exactly the desired shortest-path field. If the proof carries over to the trained model, surface reconstruction gains a principled way to handle open and non-watertight shapes with less sensitivity to noisy or sparse scans, because the method learns geometric paths rather than trusting every input point.","feed_headline":"Neural field learns shortest paths, reconstructing open surfaces","feed_subtitle":"Learning distance and gradient together yields smoother open surfaces and stronger noise tolerance.","key_machinery":"The load-bearing structure is the magnitude-direction decomposition of the network output, $F_\\theta=\\|F_\\theta\\|\\,(F_\\theta/\\|F_\\theta\\|)=d_\\theta G_\\theta$, combined with a variable-splitting constraint that makes the unit direction the gradient of the magnitude. Three loss terms drive the mechanism: the manifold loss anchors the magnitude to zero on the given points, the gradient-matching loss enforces $\\nabla d_\\theta=G_\\theta$, and the shortest-path loss requires the pulled point $x-F_\\theta(x)$ to lie on the zero level set. The shortest-path term is what forces the learned field to reproduce the identity $F=d\\nabla d$, which is the key to the global-minimum theorem, while the decomposition is what makes the $H^1$-convergence theorem for the magnitude possible.","core_discovery":"On its own terms, the central discovery is that approximating the exact shortest path $F(x)=d(x)\\nabla d(x)$ yields the distance function and its gradient simultaneously, with the distance component guaranteed to converge in $H^1(\\Omega)$. Writing the network output as $F_\\theta=d_\\theta G_\\theta$, with magnitude $d_\\theta=\\|F_\\theta\\|$ and unit direction $G_\\theta=F_\\theta/\\|F_\\theta\\|$, the loss $\\mathcal{L}_{\\mathrm{NSP}}=\\mathcal{L}_\\Gamma+\\lambda_{\\mathrm{GM}}\\mathcal{L}_{\\mathrm{GM}}+\\lambda_{\\mathrm{SP}}\\mathcal{L}_{\\mathrm{SP}}$ enforces $d_\\theta=0$ on the sampled surface, $\\nabla d_\\theta=G_\\theta$, and $d_\\theta(x-F_\\theta(x))=0$, so that points pulled by the field land on its zero level set. The paper proves that a global minimizer of this loss is an exact shortest path to a surface containing the input samples, and that the magnitude converges in the $H^1$ norm as gradient matching is driven to zero.","pith_inferences":["A testable extension is to check whether the $H^1$ guarantee survives the frozen outer gradient used in the reported shortest-path loss; proving convergence for that modified objective would require a different argument, likely a two-timescale or fixed-point analysis.","The same magnitude-direction decomposition could be applied to other unit-gradient field problems, such as geodesic distance estimation, fast marching, or level-set redistancing, where recovering distance and gradient together is equally desirable.","The minimal-area term is what separates true holes from missing scan regions; replacing it with a learned topology prior or a data-confidence mask would address the limitation the paper's conclusion identifies."],"forward_implications":["Because the loss is built around the path $x-F_\\theta(x)$ rather than around distances to the given points, the learned field is decoupled from the point cloud's quality; the paper cites this as the reason for improved robustness to noise and sparsity.","The single-network coupling of $d_\\theta$ and $G_\\theta$ yields a gradient field that is conservative by construction, avoiding the inconsistency the authors attribute to closest-surface-point methods that learn normals separately.","Convergence of $d_\\theta$ in $H^1(\\Omega)$ means the zero level set is controlled together with its gradient, reducing the kink-induced instability that appears when extracting meshes from distance functions.","The representation requires no inside or outside labels, so it applies to open and multi-layer surfaces such as garments, car interiors, and scanned indoor scenes where signed-distance or occupancy methods fail."],"supporting_citations":[{"why":"Closest prior method predicting closest surface points; NSP targets its lack of a conservative gradient field.","marker":"(Venkatesh et al., 2021)"},{"why":"Defines neural unsigned distance fields, the baseline paradigm and a comparison method for open-surface reconstruction.","marker":"(Chibane et al., 2020)"},{"why":"CAP-UDF baseline that learns unsigned distance from raw point clouds; provides the comparison and motivates robustness to corruption.","marker":"(Zhou et al., 2024)"},{"why":"Supplies the variable-splitting technique with Sobolev convergence and minimal-area regularization that NSP adapts.","marker":"(Park et al., 2024)"},{"why":"Provides the proof strategy for a similar global-minimum result in a split-PDE setup, used for Theorem 2.","marker":"(Park et al., 2023)"},{"why":"Establishes the Sobolev space $H^1(\\Omega)$ in which distance functions live and in which convergence is measured.","marker":"(Adams and Fournier, 2003)"},{"why":"Supplies the eikonal equation and viscosity-solution theory that characterize the distance function.","marker":"(Evans, 2022)"},{"why":"Dual contouring, the mesh-extraction step that turns selected surface points into the reconstructed mesh.","marker":"(Ju et al., 2002)"}],"fun_headline_variants":["Neural shortest path unifies distance and gradient for surfaces","Shortest path loss reconstructs surfaces with noise robustness","Joint distance-gradient learning via exact shortest path","Neural shortest path for robust surface reconstruction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theoretical guarantee applies to an idealized loss minimized exactly; the deployed training minimizes a modified objective with an extra minimal-area term and with one component of the shortest-path term kept fixed during optimization, and it is not proven that the modified objective has the same exact-shortest-path minimizer.","fun_headline_variants_meta":{"raw":{"variants":["Neural shortest path unifies distance and gradient for surfaces","Shortest path loss reconstructs surfaces with noise robustness","Joint distance-gradient learning via exact shortest path","Neural shortest path for robust surface reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000509,"raw_usage":{"total_tokens":2499,"prompt_tokens":983,"completion_tokens":1516,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":1456}},"tokens_in":599,"tokens_out":1516,"duration_ms":11299,"temperature":1.0,"reasoning_tokens":1456,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T16:56:44.735170+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the full network with the practical loss on a clean hemisphere point cloud, then compare the learned field against the ground-truth distance $d$ and shortest path $d\\nabla d$ on a fine grid away from the surface. If, at convergence, the $H^1(\\Omega)$ error $\\|d_\\theta-d\\|_{H^1(\\Omega)}$ or the shortest-path error $\\|F_\\theta-d\\nabla d\\|_{L^2(\\Omega\\setminus\\Gamma)}$ is bounded away from zero even as the loss approaches zero, then the reported guarantee does not transfer to the deployed training objective.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp","cited_arxiv_id":null,"evidence_quote":"Closest prior method predicting closest surface points; NSP targets its lack of a conservative gradient field."},{"cited_title":"Advances in Neural Information Processing Systems 33, 21638–21652 (2020)","cited_arxiv_id":null,"evidence_quote":"Defines neural unsigned distance fields, the baseline paradigm and a comparison method for open-surface reconstruction."},{"cited_title":"IEEE Transactions on Pattern Analysis and Machine Intelligence (2024)","cited_arxiv_id":null,"evidence_quote":"CAP-UDF baseline that learns unsigned distance from raw point clouds; provides the comparison and motivates robustness to corruption."},{"cited_title":"ReSDF: Redistancing Implicit Surfaces using Neural Networks","cited_arxiv_id":"2305.08174","evidence_quote":"Provides the proof strategy for a similar global-minimum result in a split-PDE setup, used for Theorem 2."},{"cited_title":"Elsevier (2003)","cited_arxiv_id":null,"evidence_quote":"Establishes the Sobolev space $H^1(\\Omega)$ in which distance functions live and in which convergence is measured."},{"cited_title":"American Mathematical Society 19 (2022)","cited_arxiv_id":null,"evidence_quote":"Supplies the eikonal equation and viscosity-solution theory that characterize the distance function."}],"review_version":1}