{"id":"6c0cbca2-b569-4771-938b-ca40195b0c39","arxiv_id":"2501.13963","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Maize leaves scanned by LiDAR can be converted to smooth NURBS surface models using a particle-swarm initial fit followed by a differentiable-NURBS refinement, with reported residuals of 0.02-0.07 mm.","lead":"This paper turns laser scans of corn plants into smooth, editable 3D models of the leaves using a two-stage computer optimization. If it works across many corn varieties, it would automate a tedious step in plant research and breeding.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported Chamfer distances are the training loss computed on the same points used for fitting; no held-out or ground-truth evaluation supports the 'accurate reconstruction' claim.","rationale":"The reader's weakest assumption concerns segmentation quality and the fixed 3×6 control-point topology. That is a real generality limitation, but the more load-bearing vulnerability is that the only quantitative evidence—Table 2—is an in-sample training objective. If a held-out test shows large errors, the fixed-topology limitation and the generality claim become secondary: the headline accuracy numbers would not mean what they appear to mean. The reader's rationale mentions 'in-sample Chamfer distance as the only quantitative metric without error bars or baselines,' so there is partial agreement, but it is not the reader's designated weakest assumption. The abstract also promises open-source code while the Data Availability statement conditions release on acceptance; that is a consistency problem, and the 'Smith and Doe 2024' citation with a placeholder DOI is a serious provenance red flag, but neither is the central technical load-bearing issue. The method itself is plausible: PSO initialization plus differentiable NURBS refinement is a reasonable two-stage fitting strategy, and NURBS-Diff is a published, citable module. The concern is not that the optimization cannot reduce the training Chamfer distance; it is that the paper provides no evidence this reduction corresponds to accurate reconstruction of unobserved geometry. The proposed held-out split is simple, inexpensive, and directly settles whether the reported 0.02–0.07 mm values are genuine accuracy or just training convergence. Given that the reader already rejected the paper, my analysis supports that verdict without changing it.","tokens_in":18420,"tokens_out":2989,"duration_ms":33202,"concrete_test":"For each of the 20 leaves in Table 2, randomly split the leaf point cloud into two halves: fit the PSO+NURBS-Diff pipeline using only the fit half, then compute the one-sided Chamfer distance from the held-out half to the fitted surface. Repeat with 10 random splits. If the mean held-out distance is comparable to the reported 0.02–0.07 mm (within roughly 2×), the claim survives; if it is substantially larger (e.g., >0.5 mm), then Table 2 is measuring training loss and the accuracy claim is unsupported. A Poisson surface reconstruction run on the same held-out protocol would provide a useful baseline for interpreting the numbers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests on Table 2, which reports Chamfer distances of 0.02–0.07 mm after NURBS-Diff. But the NURBS-Diff loss (Eq. 14) is a one-sided Chamfer distance between the reconstructed surface and the LiDAR point cloud, plus regularization terms; the metric tabulated in Table 2 is the same objective being minimized, evaluated on the same point cloud used for fitting. The PSO stage likewise optimizes a Chamfer/Hausdorff objective (Eq. 8) on the same cloud. Thus the table primarily documents training-set convergence, not reconstruction accuracy. Nothing is held out: there are no independent test points, no second-view scan of the same leaf, no comparison to manually measured leaf geometry, and no generic baseline such as Poisson or RBF surface reconstruction. Because the surface can pass close to fitted points while being wrong between them—especially with only 18 control points and an 8×32 evaluation grid—the reported sub-0.1 mm values are not independently meaningful. The scanner point spacing is 1.5 mm, so values below 0.1 mm require external validation rather than self-consistency. The 'across diverse genotypes' claim is supported only by visual figures; no per-genotype quantitative error is given beyond CML238 and T8. If Table 2 is just the training loss, the paper's main evidence for accurate reconstruction collapses.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage pipeline for reconstructing NURBS-based 3D models of maize leaves from LiDAR point clouds. In the first stage, particle swarm optimization (PSO) fits an initial 3×6 control-point NURBS surface using a Chamfer-plus-Hausdorff objective (Eq. 8). In the second stage, the differentiable NURBS module NURBS-Diff refines the control points with a loss that combines a one-sided Chamfer distance with curvature and proximity regularizers (Eq. 14). The authors claim that this reduces Chamfer distances to 0.02–0.07 mm (Table 2), that the method works across diverse genotypes, and that the resulting procedural models enable downstream trait extraction. The manuscript includes qualitative figures for several genotypes, runtime statistics, and an appendix with additional procedural model images.","tokens_in":18689,"tokens_out":6311,"duration_ms":62494,"significance":"If the central claim were supported by independent validation, the paper would offer a useful parametric, editable representation of maize leaves from LiDAR data, which is relevant for phenotyping and functional-structural plant modeling. The use of a differentiable NURBS module, the explicit two-stage optimization, and the stated intention to release code are positive features. However, as presented, the quantitative evidence is self-referential: the reported metric is essentially the objective minimized by the optimizer on the same point cloud, with no held-out data, no ground-truth geometry, no comparison with alternative reconstruction algorithms, and no error analysis. The cross-genotype claim rests on visual inspection and only two genotypes appear in the numerical table. A fabricated placeholder reference (\"Smith and Doe\", CropCraft) and the contradictory open-source statement further undermine the manuscript's reliability. The significance, therefore, is not yet established; the underlying idea is plausible, but the evidence is insufficient for the stated claims.","major_comments":[{"comment":"The central quantitative evidence is circular. The NURBS-Diff loss in Eq. (14) is a one-sided Chamfer distance plus regularization terms, and the 'Chamfer distance' reported in Table 2 appears to be the same objective (or a close variant) evaluated on the same point cloud used for fitting. The paper never defines the metric in Table 2, and no held-out points, independent scans, manual measurements, or alternative reconstruction baselines are provided. Thus the reduction from ~0.1–1.0 mm to 0.02–0.07 mm mostly documents convergence of the training loss, not reconstruction accuracy. This is especially troubling because the scanner point spacing is 1.5 mm, so sub-0.1 mm values require external validation. The abstract's claim of 'accurate 3D reconstruction' and the conclusion's claim that the method 'greatly improves the quality of fit' are therefore not supported by the evidence as presented.","section":"Section 3.3, Eq. (14) and Table 2"},{"comment":"The claim of 'diverse genotypes' is not quantitatively supported. Table 2 reports numerical results for only two genotypes (CML238 and T8), ten leaves each, with no error bars, no standard deviations, and no repetition of the stochastic PSO runs. The other genotypes (M162W, CI90C, B73, Mo17, etc.) appear only in qualitative figures. The fixed 3×6 control-point topology and parameter bounds in Section 3.2 may not capture leaves with complex double curvature or twisting, yet no analysis is provided to show when this topology fails. The claim of robustness and adaptability across genotypes therefore goes beyond the evidence.","section":"Section 4, Table 2 and Figures 7–12"},{"comment":"The citation 'John Smith and Jane Doe, Cropcraft: Inverse procedural modeling for realistic 3D crop canopies, Proceedings of the ACM SIGGRAPH Conference, 2024, doi: 10.1145/1234567.8901234' in Section 1 and in the reference list is a placeholder/fabricated reference; the DOI has no real registry. This citation must be removed or replaced with a legitimate source. In addition, the abstract states 'All our codes are open-source,' but the Data Availability statement says the code 'will be publicly available upon acceptance,' and no repository link is provided. These are integrity issues that must be corrected.","section":"References and Data Availability"},{"comment":"No comparison with existing surface reconstruction or NURBS-fitting methods is provided, even though the paper cites relevant work (e.g., Gálvez and Iglesias 2012 for PSO-based NURBS fitting, Kazhdan and Hoppe 2013 for Poisson reconstruction). Without a baseline, the claimed improvement in fit is not contextualized, and the reader cannot judge whether the two-stage method offers an advantage over simpler or cheaper alternatives. This is a load-bearing omission for the paper's contribution claim.","section":"Section 4 and Table 3"}],"minor_comments":[{"comment":"The 'one-sided Chamfer distance' used in the loss is not defined in the manuscript. Please provide the exact formula and clarify how it differs from the two-sided version in Eq. (9).","section":"Section 3.3, Eq. (14)"},{"comment":"Please specify how many plants and which genotypes were used to compute the mean and standard deviation in Table 3, and clarify whether Figure 10 includes all 200 plants or only a subset with reported metadata.","section":"Section 4, Table 3 and Figure 10"},{"comment":"The sentence 'also see Table 1 that shows the procedural model side by side with the 2D photos' cross-references the wrong table: Table 1 in the main text is the PSO hyperparameter table, while the procedural models appear in the Appendix tables that are also numbered Table 1 and Table 2. Please renumber the appendix tables and fix the reference.","section":"Section 4, paragraph after Figure 9"},{"comment":"The conclusion mentions 'dealing with missing points in the point cloud data' as a main challenge and claims the optimization reduces their influence, but no quantitative robustness study or ablation on missing-data fraction is presented. Either add such an experiment or temper the claim.","section":"Section 5"},{"comment":"The appendix reuses the table numbers 'Table 1' and 'Table 2' from the main text, which causes confusion. Please use unique labels such as Table A.1 and Table A.2.","section":"Appendix"}],"recommendation":"reject","confidential_remarks":"The fabricated 'Smith and Doe' reference is a serious scholarly-integrity concern that the editor should investigate. Beyond that, the paper's main evidence is a self-evaluated training objective; the authors would need to provide independent validation, baseline comparisons, and honest genotype-by-genotype reporting to make the central claim credible. The open-source statement also needs to be reconciled with the actual code availability. These issues, in my view, go beyond a routine revision and justify rejection at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this one. The pipeline—PSO for a coarse NURBS fit, then NURBS-Diff for refinement—is a sensible two-stage approach to converting maize leaf LiDAR point clouds into CAD-style surfaces, and it is demonstrated on real field-grown plants across several genotypes. The component pieces are known (PSO-NURBS fitting goes back to Gálvez and Iglesias; NURBS-Diff is the same group's prior module), so the new contribution is the application and the specific coupling. That is a legitimate but modest extension.\n\nThe larger issue is that the paper's main quantitative evidence for \"accurate reconstruction\" does not actually measure accuracy. Table 2 reports Chamfer distances between the fitted surface and the same point cloud used for fitting; it is the training loss, not an independent evaluation. There are no held-out points, no ground-truth measurements, no second scans, and no baseline such as Poisson surface reconstruction. Given the scanner's 1.5 mm point spacing, sub-0.1 mm values computed on in-sample points are not meaningful as evidence of fidelity. The qualitative figures look plausible, but that is not the same as a validated reconstruction.\n\nThere are also two concrete red flags. The citation to Smith and Doe 2024 (CropCraft) with the DOI 10.1145/1234567.8901234 is clearly a placeholder—this should not have survived any check, and the authors need to either provide the real reference or remove it. And the abstract promises \"all our codes are open-source\" while the Data Availability section says code will be public upon acceptance; that mismatch should be fixed. Minor additional gaps: no error bars in Table 2, only two genotypes have quantitative results despite the cross-genotype claim, and the phyllotaxy extraction is described but never evaluated quantitatively.\n\nThe fixed 3×6 control-point topology may limit how well the surface can represent strong double curvature or twisted leaves; that is a testable limitation rather than a fatal flaw.\n\nWho is this for? The plant phenotyping / FSPM crowd, and anyone who wants an editable, CAD-friendly bridge from point clouds to leaf models. I would send it to a serious referee, not desk-reject it, because the core idea is practical and the evaluation gaps are fixable. The referee should ask for code and data release, a real evaluation protocol (held-out or manually measured geometry, plus at least one standard reconstruction baseline), and a correction of the citation issue. I would not cite this version in my own work, but I would look at a revised version.","headline":"A practical two-stage NURBS fitting pipeline for maize leaves, but the quantitative evidence is in-sample training error and the paper carries a fabricated citation; worth a serious referee only with major revisions.","tokens_in":19303,"tokens_out":3997,"would_cite":false,"duration_ms":39429,"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":"Two-stage optimizer turns maize LiDAR scans into editable 3D leaf models","keywords":["procedural modeling","NURBS surfaces","LiDAR point clouds","3D plant phenotyping","particle swarm optimization","differentiable programming","maize architecture","phyllotaxy"],"falsifier":"Take a maize leaf with pronounced twisting or double curvature, segment it cleanly, and run the pipeline; if the final Chamfer distance does not fall below roughly 0.1 mm or the fitted surface visibly flattens the twist, the fixed 3×6 NURBS topology cannot represent such leaves and the claimed cross-genotype generality fails.","tokens_in":18220,"feed_emoji":"🌽","tokens_out":5945,"duration_ms":52007,"temperature":0.7,"pith_summary":"This paper claims that a two-stage optimization pipeline can automatically convert raw LiDAR point clouds of field-grown maize plants into editable, CAD-style 3D leaf models, without manual modeling, across diverse genotypes. The first stage uses particle swarm optimization (PSO) to fit a coarse NURBS surface to each segmented leaf; the second stage refines it with a differentiable NURBS module, reducing the Chamfer distance from roughly 0.1–1.0 mm to 0.02–0.07 mm. If the claim holds, plant scientists gain a scalable route from unstructured scans to structured parametric models from which traits such as leaf angle, curvature, and phyllotaxy can be read directly, and the fitted surfaces can serve as digital twins for simulation. The paper demonstrates the pipeline on multiple maize inbred lines and reports that a full ten-leaf plant reconstructs in about an hour.","feed_headline":"Two-step fit turns maize LiDAR scans into editable 3D leaf models","feed_subtitle":"PSO plus differentiable NURBS surfaces cut fit error to tens of microns, enabling trait extraction.","key_machinery":"The central object is a bi-cubic NURBS surface with a fixed $3\\times 6$ grid of control points, defined by 32 parameters: six shared $x$-coordinates, six mid-row $y$-values plus six $\\Delta y$ offsets that create symmetric tapering, and fourteen $z$-values that capture curvature. A Non-Uniform Rational B-Spline (NURBS) surface is a standard CAD representation defined by control points, weights, and knot vectors; here the weights and knot vectors are held constant. The argument is carried by the two-stage optimizer: PSO provides a global initial fit using a loss that combines Chamfer and Hausdorff distances, and NURBS-Diff refines the control points with an Adam optimizer on an $8\\times 32$ evaluation grid, using a loss that adds curvature and proximity penalties to the Chamfer distance.","core_discovery":"The central discovery is that the combination of global PSO initialization and gradient-based differentiable NURBS refinement yields accurate 3D reconstructions of maize leaves from LiDAR point clouds, with fidelity sufficient for downstream phenotyping. The authors show that PSO alone captures the overall leaf shape but leaves systematic error near edges and tips, while the subsequent refinement lowers the Chamfer distance by roughly an order of magnitude across all reported leaves. The fixed $3\\times 6$ bi-cubic NURBS control-point topology, with shared $x$-coordinates and symmetric $y$ offsets, encodes the prior that a maize leaf is a single smoothly tapering, curved surface, and the optimization adjusts 32 parameters (plus one rotation angle) to fit it to each leaf's point cloud.","pith_inferences":["The accuracy numbers are fit-to-scan errors, not ground-truth leaf geometry; an external validation against manual leaf measurements would be needed to confirm that the reconstructed surfaces are phenotypically accurate.","Because the fixed $3\\times 6$ topology assumes a single smooth tapered surface, leaves with strong twisting or double curvature would likely break the fit; testing on such leaves would delineate the true scope of the cross-genotype claim.","The one-hour runtime and the manual or visual segmentation step suggest that the highest-leverage improvement is replacing the PSO initialization with a learned predictor and automating leaf separation, which would make the pipeline truly high-throughput.","The same two-stage recipe of meta-heuristic initialization followed by differentiable NURBS refinement could transfer to other organs or species whose leaves are roughly single-sheet surfaces, provided the parameterization is adjusted."],"forward_implications":["Reported Chamfer distances fall from 0.09–1.02 mm after PSO to 0.02–0.07 mm after NURBS-Diff, a roughly tenfold improvement in fit fidelity.","A complete ten-leaf maize plant reconstructs in about one hour ($3430 \\pm 490$ s on a 96-core CPU), with each leaf fit independently and therefore parallelizable.","Because the output is a parametric NURBS surface, traits such as leaf length, width, curvature, and phyllotaxy can be computed from the fitted control points rather than from raw point clouds.","The curvature and proximity penalties in the loss let the method bridge gaps in incomplete point clouds, demonstrated on a T8 leaf with missing data.","The pipeline is demonstrated on diverse SAM-panel genotypes including B73, Mo17, CML238, T8, M162W, and CI90C, suggesting no per-genotype retuning is required."],"supporting_citations":[{"why":"Provides the differentiable NURBS module (NURBS-Diff) used for the gradient-based refinement stage.","marker":"Prasad et al., 2022"},{"why":"Establishes particle swarm optimization as a method for fitting NURBS surfaces to point clouds, which the first stage builds on.","marker":"Gálvez and Iglesias, 2012"},{"why":"Supplies the SAM diversity panel of maize genotypes used to test cross-genotype generalization.","marker":"Thompson et al., 2015"},{"why":"Characterizes morphometric diversity in the maize shoot apical meristem, grounding the claim that the panel spans diverse leaf architectures.","marker":"Leiboff et al., 2015"}],"fun_headline_variants":["Two-step fit turns LiDAR into 3D maize leaf models","PSO plus NURBS-Diff: accurate maize leaf reconstruction","Maize phenotyping via differentiable NURBS and PSO","From point clouds to editable maize: hybrid NURBS pipeline"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes each leaf has already been separated from the stalk and other leaves in the point cloud, and that every maize leaf can be represented by a single bi-cubic NURBS surface with the fixed 3×6 control-point layout and parameter bounds; leaves with complex double curvature, twisting, or segmentation errors break that assumption.","fun_headline_variants_meta":{"raw":{"variants":["Two-step fit turns LiDAR into 3D maize leaf models","PSO plus NURBS-Diff: accurate maize leaf reconstruction","Maize phenotyping via differentiable NURBS and PSO","From point clouds to editable maize: hybrid NURBS pipeline"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00053,"raw_usage":{"total_tokens":2548,"prompt_tokens":933,"completion_tokens":1615,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":1542}},"tokens_in":549,"tokens_out":1615,"duration_ms":11652,"temperature":1.0,"reasoning_tokens":1542,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:06:06.943910+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a maize leaf with pronounced twisting or double curvature, segment it cleanly, and run the pipeline; if the final Chamfer distance does not fall below roughly 0.1 mm or the fitted surface visibly flattens the twist, the fixed 3×6 NURBS topology cannot represent such leaves and the claimed cross-genotype generality fails.","supporting_citations":[{"cited_title":"Particle swarm optimization for non-uniform rational B-Spline (NURBS) surface reconstruction from clouds of 3D data points","cited_arxiv_id":null,"evidence_quote":"Establishes particle swarm optimization as a method for fitting NURBS surfaces to point clouds, which the first stage builds on."},{"cited_title":"Diversity of maize shoot apical meristem architecture and its relationship to plant morphology","cited_arxiv_id":null,"evidence_quote":"Supplies the SAM diversity panel of maize genotypes used to test cross-genotype generalization."},{"cited_title":"Schnable, and Michael J","cited_arxiv_id":null,"evidence_quote":"Characterizes morphometric diversity in the maize shoot apical meristem, grounding the claim that the panel spans diverse leaf architectures."}],"review_version":1}