{"id":"780d126f-9e26-4df7-8f8e-36596a91f38d","arxiv_id":"2507.12714","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A neural parametric model for leaves that disentangles 2D base shape from 3D deformation, learned from 2D image data plus a new 300-pair 3D scan dataset, and fitted to observations for reconstruction.","lead":"NeuraLeaf builds a neural model of 3D leaves that separates each leaf's flat outline from its 3D bending, so the outline can be learned from ordinary 2D images. It also adds a dataset of roughly 300 scanned leaf pairs and shows the model can fit new depth or point-cloud observations more accurately than earlier leaf, spline, and neural baselines.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 15.1→2.1 mm Chamfer gain over NPM may be an artifact of the correspondence-based NPM baseline and biased DeformLeaf registrations, not of NeuraLeaf's representation.","rationale":"The central claim requires both a genuinely generalizable shape/deformation space and a fair, accurate evaluation. The least secure condition is the evaluation: Table 1 is the only head-to-head evidence for the headline 2.1 vs 15.1 mm improvement. Both numbers are threatened by the combination of a correspondence-supervised NPM baseline and a registration pipeline whose accuracy is never validated. This overlaps with the reader's weakest_assumption about DeformLeaf ground truth, but the specific asymmetry makes the concern sharper: even if the Metashape geometry were unbiased, noisy ARAP/CPD correspondences could alone create the observed gap by corrupting only the NPM training signal. The proposed concrete tests settle the point: a correspondence-free NPM baseline removes the confound, and a synthetic registration benchmark quantifies the ground-truth error. Because the paper otherwise documents the method carefully and releases code and data, I do not see the central construction as invalid; rather, the existing conditional acceptance is the right state, with these checks as the explicit conditions.","tokens_in":17141,"tokens_out":9253,"duration_ms":124275,"concrete_test":"Retrain the NPM baseline with a correspondence-free objective (Chamfer plus the same length/Laplacian regularizers used for NeuraLeaf) and recompute Table 1; if its Chamfer distance drops substantially from 15.1 mm, the comparison is confounded. Independently, synthetically deform leaf meshes with known ground truth, render multi-view images, run Metashape plus the ARAP/CPD pipeline, and compare the recovered surfaces and correspondences against the known deformations; if registration error grows with curvature or exceeds roughly 1 mm on strongly curved leaves, the DeformLeaf ground truth cannot support the reported 2.1 mm claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The main quantitative evidence (Table 1) compares NeuraLeaf to an NPM baseline that, per Suppl. B.3, is trained with the dense correspondences stored in DeformLeaf. Those correspondences are produced by ARAP followed by CPD (Sec. 4 / Suppl. A.2-A.3). NeuraLeaf's deformation training is deliberately correspondence-free and minimizes Chamfer distance to the raw 3D scans (Sec. 3.3, Eq. 11). If the ARAP/CPD registrations are inaccurate for strongly curved or occluded leaves, the NPM baseline is trained with corrupted displacement targets, while NeuraLeaf is not penalized by those errors. The reported 7x Chamfer improvement could therefore measure robustness to registration noise rather than superiority of the disentangled representation. The evaluation compounds this: both the training signal and the test metric come from the same Metashape reconstruction pipeline, so systematic bias in that pipeline for exactly the strong deformations the method claims to model would propagate into the learned deformation space and the 2.1 mm number. The paper's limitation section acknowledges limited deformation variety but does not evaluate registration accuracy or baseline fairness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes NeuraLeaf, a neural parametric model for 3D leaves that separates leaf geometry into a 2D base shape (learned from 2D leaf silhouette datasets) and a 3D deformation (learned from a newly captured DeformLeaf dataset of paired flattened and deformed leaves). Deformation is modeled by a skeleton-free linear blend skinning scheme with up to 1,000 control points, where a shape-conditioned decoder predicts skinning weights and a deformation-conditioned decoder predicts rigid transformations. The paper further contributes a correspondence-free training objective for deformation, a regularizer that links the deformation latent norm to Chamfer magnitude, and latent-code inversion for fitting to depth/point-cloud observations. Experiments on single-leaf and multi-leaf reconstruction compare NeuraLeaf with PCA, B-spline, and neural parametric model (NPM) baselines, reporting substantially lower Chamfer distances, plus qualitative results for generation, interpolation, deformation transfer, and comparison with image/text-to-3D methods.","tokens_in":17452,"tokens_out":8553,"duration_ms":91382,"significance":"NeuraLeaf is a sensible and potentially useful representation for leaf modeling: the disentanglement leverages abundant 2D leaf data, the skeleton-free skinning avoids the need for a common leaf skeleton, and the authors release both code and a new dataset (DeformLeaf) with dense correspondences. The correspondence-free training is a genuine advantage over displacement-field NPMs, and the qualitative deformation-transfer results support the plausibility of a shared deformation space. However, the strength of the contribution is not yet matched by the evidence: the quantitative claims rest on a small test set with no error bars, and the main baseline comparison may be confounded by the very registration pipeline the baseline depends on, so the current evaluation cannot fully support the magnitude of the reported improvement.","major_comments":[{"comment":"Equation (7) is not a valid linear blend skinning formula as written. In homogeneous coordinates, tilde{v}'_i - tilde{c}_k = [v'_i - c_k, 0]^T, so applying T_k and taking the weighted sum yields a vector with zero homogeneous component (a direction), not a point in 3D. The standard formulation is tilde{v}_i = sum_k w_{i,k} T_k tilde{v}'_i, or, for transformations about control points, tilde{v}_i = sum_k w_{i,k}(T_k(tilde{v}'_i - tilde{c}_k) + tilde{c}_k). Please correct the equation and confirm that the implemented code uses the corrected form; because Eq. (7) defines the entire deformation model, this is a load-bearing issue.","section":"§3.3, Eq. (7)"},{"comment":"Table 1 reports a single aggregate Chamfer distance over 30 test leaves, and Table 2 reports results on 5 plants, with no error bars, confidence intervals, or significance tests. Given the large claimed improvement (15.1 → 2.1 mm), the authors should report per-leaf/per-plant distributions (or at least standard deviations) and a paired test, and ideally a per-species breakdown, to show that the gain is systematic rather than driven by outliers. Without this, the headline quantitative claim is not statistically grounded.","section":"§6.2, Tables 1 and 2"},{"comment":"The comparison with the NPM baseline is potentially unfair because the baseline is trained using the dense correspondences in DeformLeaf (Suppl. B.3), which are produced by the ARAP+CPD pipeline (Sec. 4, Suppl. A.2–A.3), whereas NeuraLeaf is trained with a correspondence-free Chamfer loss (Eq. 11). If ARAP/CPD alignments are inaccurate for strongly curved or occluded leaves, the NPM baseline is penalized by corrupted displacement targets, while NeuraLeaf is not. The paper should quantify registration accuracy (e.g., against manual landmarks) on the test set, and/or train an NPM variant with a fitting objective that does not rely on those correspondences, to demonstrate that the 7× improvement reflects the representation rather than registration noise. Relatedly, because training and test deformed shapes both come from the same Metashape reconstruction pipeline, an evaluation on independent data (e.g., synthetic leaves with known ground-truth deformation) would help rule out that the reported numbers are tied to the specific reconstruction biases of that pipeline.","section":"Suppl. B.3 vs. §3.3–§4"},{"comment":"Equation (10) defines Lmap = (dcham(Sb,Sd)/||zd||2 − phi)^2, but the Chamfer distance between the base and deformed shape is not a pure deformation-magnitude measure: for the same physical deformation, a larger or differently shaped leaf yields a different dcham. The regularizer therefore couples the deformation latent norm to base-shape geometry, which may weaken the claimed disentanglement between z_s and z_d. The paper should analyze this coupling and ensure the denominator is safely away from zero (e.g., add a small epsilon or a lower bound on ||zd||2).","section":"§3.3, Eq. (10)"}],"minor_comments":[{"comment":"The loss term L_sil is used in the training objective but is never defined in the text; please define it.","section":"§3.1, Eq. (5)"},{"comment":"The meaning of the columns 'Corres-free' and 'Temp-free' is not explained; please state explicitly whether each method requires dense correspondences or a template.","section":"§6.1, Table 1 caption"},{"comment":"The text says DeformLeaf contains approximately 300 base-deformation pairs, while the figure caption reads '≈ 300 species & 4000 instances'; please disambiguate which quantity refers to DeformLeaf and which to the 2D dataset.","section":"§4 and Fig. 2"},{"comment":"The PCA baseline reports a Chamfer distance of 1311.3 mm, an order of magnitude larger than the other entries; please check for a misalignment or typo and, if real, discuss why it is so large.","section":"§6.2, Table 2"},{"comment":"The texture generator is denoted f_theta_t in the text but the figure uses 'E/D'; please unify the notation.","section":"§3.2 and Fig. 2"},{"comment":"The multi-leaf reconstruction section says the shape latent is shared for observations of the same plant species, but the clustering and anchor details are only in Suppl. C.4; consider moving the essential mechanism (K-means, anchor regularization) to the main text or giving a one-sentence summary, since it is important for interpreting Table 2.","section":"§6.2"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper is a solid conference-level contribution with a useful dataset release, but the journal version needs stronger statistical evaluation and an analysis of registration-induced bias in the baseline comparison. I see no grounds for concern about novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"NeuraLeaf is the first neural parametric model for leaves, and it is a genuinely useful piece of engineering. The disentanglement of a 2D base shape (learnable from large 2D leaf image sets) from a 3D deformation space is a clean idea, and the skeleton-free LBS with 1000 control points is a sensible departure from articulated-object NPMs. The DeformLeaf dataset, with roughly 300 base-deformed pairs, is a real contribution even if the paper around it is modest. The code and data are public, the ablations are informative (control-point count, latent inversion, boundary losses), and the qualitative results look right. Credit is earned.\n\nThe soft spots are mostly about the evaluation, not the method. The headline number is 2.1 mm Chamfer versus 15.1 for the NPM baseline. But that baseline is trained with the dense correspondences stored in DeformLeaf, which come from an ARAP+CPD pipeline whose accuracy is never assessed. NeuraLeaf is trained correspondence-free, optimizing Chamfer to the raw scans, and evaluated with the same Chamfer metric. So part of the gains could be robustness to registration noise rather than a better representation. This is a legitimate concern, and the paper should address it by reporting an NPM baseline trained with the same Chamfer loss, or by evaluating registration accuracy on a subset. It is not a fatal flaw: NeuraLeaf is still the only method here that does not require correspondences, and the disentanglement is novel.\n\nThe evaluation scale is small: 30 test leaves, 5 plants, no error bars. The test set comes from the same DeformLeaf acquisition pipeline, so systematic bias in Metashape reconstructions of strongly curved leaves would propagate into both training and evaluation. A per-species breakdown and some measure of reconstruction reproducibility would strengthen the claims. The limitations paragraph in the paper is honest: deformation variety is limited, and texture fitting is not solved. That is acceptable for a first model.\n\nWho this is for: people building plant reconstruction or foliage modeling pipelines, and researchers working on neural parametric representations for non-articulated deformable objects. It deserves a serious referee. I would send it out, but with instructions to the authors to add a fairer baseline and error bars before acceptance.","headline":"A solid first neural parametric leaf model with a valuable dataset; the quantitative edge over NPM is real but overstated by an unfair baseline and small test set.","tokens_in":17952,"tokens_out":2792,"would_cite":true,"duration_ms":32696,"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":"NeuraLeaf models each leaf as a flat 2D base shape plus a learned 3D deformation, and fits partial 3D observations of leaves with much lower Chamfer distance than existing parametric leaf models.","keywords":["neural parametric model","leaf reconstruction","3D deformation modeling","skeleton-free skinning","linear blend skinning","shape and deformation disentanglement","DeformLeaf dataset","signed distance function"],"falsifier":"Take a set of strongly curled leaves, scan them with an independent high-resolution laser or structured-light scanner, fit NeuraLeaf to those scans, and compare against the PCA, B-spline, and NPM baselines; if the Chamfer advantage over the NPM baseline shrinks to noise, or if deformation codes no longer order by actual bending amount, the claim that the learned deformation space is accurate is refuted.","tokens_in":16995,"feed_emoji":"🍃","tokens_out":9610,"duration_ms":90092,"temperature":0.7,"pith_summary":"NeuraLeaf is a neural parametric model built around the observation that a flattened leaf is nearly a 2D plane. The paper argues that splitting leaf geometry into a 2D base shape and a 3D deformation makes it possible to learn realistic leaf shape from abundant 2D leaf-image datasets while learning deformation from a comparatively small 3D dataset. To support the deformation side, the authors introduce a skeleton-free skinning model with up to a thousand control points and a new captured dataset, DeformLeaf, of roughly 300 base-to-deformed leaf pairs. On the 30-leaf DeformLeaf test set, fitting NeuraLeaf to single-view point clouds reaches a Chamfer distance of 2.1 mm and normal consistency of 0.973, versus 15.1 mm and 0.961 for the neural parametric baseline. If this stands, 3D leaf reconstruction and generation become practical from inputs as simple as depth maps and point clouds.","feed_headline":"NeuraLeaf cuts leaf-reconstruction error from 15.1 to 2.1 mm","feed_subtitle":"A flat base shape plus a shared deformation space lets the model learn from 2D photos and fit real 3D scans.","key_machinery":"The load-bearing object is the skeleton-free neural blend skinning model. Instead of a skeletal joint hierarchy, NeuraLeaf places up to 1,000 control points uniformly over the leaf's UV plane; the skinning weight decoder maps each base-shape vertex, conditioned on the shape latent code, to a distribution over control points, and the transformation decoder maps each control point, conditioned only on the deformation latent code, to a rigid transformation. Linear blend skinning then deforms the base mesh. Because leaves have no common skeleton and deform continuously, the many-control-point design is what lets the model capture fine surface details such as venation while keeping the deformation latent space independent of species. The accompanying deformation mapping loss ties the norm of the deformation code to the chamfer distance between base and deformed shape, which gives the latent space a meaningful zero-mean Gaussian structure.","core_discovery":"The paper's central claim is that a leaf's full 3D geometry can be faithfully represented by two independent latent codes: a shape latent code that controls the flattened base shape, and a deformation latent code that controls how that base mesh bends in 3D. The base shape is a signed distance function in the plane—each point records its signed distance to the leaf outline—converted to a base mesh, so the shape space can be trained on 2D leaf-image data. The deformation is produced by a skeleton-free linear blend skinning model: a dense set of up to 1,000 control points on the base plane, with skinning weights predicted from the shape latent and rigid transformations (quaternion plus translation) predicted from the deformation latent alone. Because the deformation space is shared across all base shapes, the same deformation code applies to different leaf species, and fitting an observed point cloud or depth map amounts to inverting the two latent codes and then refining them by direct optimization. The paper reports that this representation, trained with a chamfer-based deformation mapping loss on the new DeformLeaf dataset, reconstructs test leaves with substantially lower Chamfer distance than PCA, B-spline, and human-oriented neural parametric baselines.","pith_inferences":["If the disentanglement holds, the deformation space should transfer not only to unseen species but to synthetic or edited base shapes; a cheap test is to deform procedurally generated leaves with latent codes fitted to real scans and check physical plausibility.","The paper's own reported limitation that rare deformation types are undercovered suggests the bottleneck is dataset species diversity rather than representation; adding scans of strongly curled, lobed, or needle-like leaves would show whether the 1,000-point skinning model scales.","The reported accuracy depends on photogrammetric ground truth registered by ARAP and CPD; an independent evaluation against laser scans would separate representation quality from dataset bias.","Because the transformation decoder is conditioned only on the deformation code, the model implicitly assumes a canonical alignment between base and deformed shapes; future work might test whether the learned space remains coherent when leaves are observed from arbitrary initial poses."],"forward_implications":["Because base shapes are 2D, the shape space can be trained on large existing collections of 2D leaf scans, sidestepping the scarcity of 3D leaf data.","Because the deformation code is shared across species, a deformation learned on one leaf can be transferred to another base shape, giving a simple route to generating deformed foliage in computer graphics.","Reconstruction from partial observations reduces to latent-code inversion followed by direct refinement, so depth maps and point clouds from RGB-D cameras can be turned into complete leaf models.","Texture is generated as a UV map aligned with the base shape, so appearance stays consistent with geometry during deformation.","When reconstructing multiple occluded leaves from a top-view RGB-D image, sharing the shape latent across instances recovers leaf areas hidden by occlusion."],"supporting_citations":[{"why":"Defines the neural parametric model paradigm that NeuraLeaf extends to leaves and uses as the main quantitative baseline in reconstruction.","marker":"[33]"},{"why":"Supplies the linear blend skinning formula that deforms base-mesh vertices from control-point transformations.","marker":"[17]"},{"why":"Provides the large 2D leaf image repository used to train the base shape space.","marker":"[9]"},{"why":"Supplies the image-to-image translation architecture used for the texture generator that maps a mask to a UV texture.","marker":"[58]"},{"why":"Provides As-Rigid-As-Possible registration, the first stage of establishing dense correspondences in DeformLeaf.","marker":"[46]"},{"why":"Provides Coherent Point Drift non-rigid registration, the second stage that aligns strongly deformed leaf pairs.","marker":"[10]"},{"why":"Provides the PCA-based parametric foliage model used as a baseline in single- and multiple-leaf reconstruction.","marker":"[5]"},{"why":"Provides the B-spline curve representation used as a baseline parametric leaf model.","marker":"[12]"}],"fun_headline_variants":["NeuraLeaf disentangles leaf shape and deformation for 3D reconstruction","Leaf 3D modeling: NeuraLeaf learns from 2D photos and fits 3D scans","Skeleton-free leaf skinning: NeuraLeaf's shape-deformation split","New leaf dataset DeformLeaf powers NeuraLeaf's dual-latent model","NeuraLeaf: flat shape plus shared deformation reconstructs leaves"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported accuracy rests on the assumption that the DeformLeaf ground truth—photogrammetric 3D scans of deformed leaves aligned to flattened 2D photos through rigid and non-rigid registration—is unbiased; if that alignment is wrong for strongly curved leaves, both the learned deformation space and the reported Chamfer improvements inherit the error.","fun_headline_variants_meta":{"raw":{"variants":["NeuraLeaf disentangles leaf shape and deformation for 3D reconstruction","Leaf 3D modeling: NeuraLeaf learns from 2D photos and fits 3D scans","Skeleton-free leaf skinning: NeuraLeaf's shape-deformation split","New leaf dataset DeformLeaf powers NeuraLeaf's dual-latent model","NeuraLeaf: flat shape plus shared deformation reconstructs leaves"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000874,"raw_usage":{"total_tokens":3810,"prompt_tokens":1001,"completion_tokens":2809,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":2705}},"tokens_in":617,"tokens_out":2809,"duration_ms":22249,"temperature":1.0,"reasoning_tokens":2705,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:40:28.548182+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of strongly curled leaves, scan them with an independent high-resolution laser or structured-light scanner, fit NeuraLeaf to those scans, and compare against the PCA, B-spline, and NPM baselines; if the Chamfer advantage over the NPM baseline shrinks to noise, or if deformation codes no longer order by actual bending amount, the claim that the learned deformation space is accurate is refuted.","supporting_citations":[{"cited_title":"NPMs: Neural parametric models for 3D deformable shapes","cited_arxiv_id":null,"evidence_quote":"Defines the neural parametric model paradigm that NeuraLeaf extends to leaves and uses as the main quantitative baseline in reconstruction."},{"cited_title":"Skinning: Real-time shape deformation","cited_arxiv_id":null,"evidence_quote":"Supplies the linear blend skinning formula that deforms base-mesh vertices from control-point transformations."},{"cited_title":"A data repository of leaf images: Practice towards plant conservation with plant pathology","cited_arxiv_id":null,"evidence_quote":"Provides the large 2D leaf image repository used to train the base shape space."},{"cited_title":"Unpaired image-to-image translation using cycle- consistent adversarial networks","cited_arxiv_id":null,"evidence_quote":"Supplies the image-to-image translation architecture used for the texture generator that maps a mask to a UV texture."},{"cited_title":"As-rigid-as-possible surface modeling","cited_arxiv_id":null,"evidence_quote":"Provides As-Rigid-As-Possible registration, the first stage of establishing dense correspondences in DeformLeaf."},{"cited_title":"Coherent point drift revisited for non-rigid shape match- ing and registration","cited_arxiv_id":null,"evidence_quote":"Provides Coherent Point Drift non-rigid registration, the second stage that aligns strongly deformed leaf pairs."},{"cited_title":"Image-based reconstruction and synthesis of dense foliage","cited_arxiv_id":null,"evidence_quote":"Provides the PCA-based parametric foliage model used as a baseline in single- and multiple-leaf reconstruction."},{"cited_title":"Gordon and Richard F","cited_arxiv_id":null,"evidence_quote":"Provides the B-spline curve representation used as a baseline parametric leaf model."}],"review_version":1}