{"id":"72800ee5-ee95-4e6e-8483-40bb348d4101","arxiv_id":"2507.07521","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Cubic Hermite spline trajectories with analytic velocity and acceleration regularizers improve sparse-frame interpolation and spatial coherence in dynamic scene reconstruction.","lead":"This paper proposes a spline-based representation for deforming 3D scenes, where motion trajectories are modeled with cubic Hermite splines whose knots and tangents are predicted by a neural network. It shows better temporal interpolation from sparse frames and smoother spatial motion than prior deformation-field methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported Moran's I values do not follow Moran's I as defined: Eq. 12 omits mean-centering and uses raw offsets, so the spatial-coherence evidence needs rechecking.","rationale":"I read the paper as making two related claims: explicit spline interpolation improves temporal interpolation for sparse signals, and it improves spatial coherence in dynamic scene reconstruction. The first claim is supported by the EPE numbers in Table 1 and is largely independent of the knot-count concern; even if N is a tuned hyperparameter, the comparison against ResFields still shows a consistent EPE advantage, and the ablations on L_v and L_acc indicate that the spline interpolation itself, not just the regularizers, drives the gain. The reader's weakest_assumption about N = T/2 is valid and is partly self-acknowledged in Appendix D, where the authors note that dense knots increase jitter and where Sec. 3.2.1 manually sets N = 30 on NeRF-DS. That concern affects the paper's claim of a principled, heuristic-free knot selection, but it does not by itself invalidate the main EPE comparison. The more load-bearing issue is the Moran's I metric. The formula printed in Eq. 12 is not a version of Moran's I: it omits mean centering, uses raw motion vectors in the numerator, and uses a non-squared raw sum in the denominator. Because the paper uses M.'s I as the headline quantitative evidence for improved spatial coherence across all dynamic-scene-reconstruction tables, a misspecified index undermines precisely the part of the central claim that is not already supported by photometric metrics. The fix is straightforward: implement the standard centered statistic, rerun the tables, and check whether the rankings survive. If they do, the concern is mostly terminological; if they do not, the coherence claim would need to be substantially downgraded. For that reason the reader's CONDITIONAL verdict remains appropriate: the EPE-based core is credible and independently testable, but the spatial-coherence evidence must be corrected and re-reported before the broader claim can be accepted.","tokens_in":22883,"tokens_out":13704,"duration_ms":164196,"concrete_test":"Recompute the M.'s I columns of Tables 1-4 using the standard centered Moran's I (or its local form) with the same K-nearest-neighbor weights w_ij but with centered vectors z_i = v_i - mean(v), denominator sum_i ||z_i||^2, and factor N/W for global I (or K/sum_w for local I). Also run the null check: randomly permute v_i across points and verify that the corrected I drops to near zero, whereas Eq. 12 as printed does not. If the rank ordering of SF variants versus ResFields, 4DGS, and Grid4D persists after the correction, the spatial-coherence claim survives; if not, it loses its current quantitative support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section A.1's Eq. 12 is not Moran's I. The standard statistic is I = (N/W) * [sum_i sum_j w_ij (z_i - zbar)(z_j - zbar)] / [sum_i (z_i - zbar)^2]. Eq. 12 instead computes [K/sum_w] * [sum_i sum_j w_ij v_i·v_j] / [sum_i v_i] with v_i = x_i(t+1) - x_i(t) as 3D offsets and no centering; the denominator is neither squared nor centered, and for vector-valued v_i the printed denominator is dimensionally inconsistent. An uncentered dot-product index is dominated by the mean motion (e.g., global translation or camera motion) and can report high autocorrelation even for spatially random velocities superimposed on a common offset. Since the paper's claimed improvement in spatial coherence in Tables 1-4 rests on the M.'s I column, and the metric is introduced as a 'principled and robust evaluation,' this is not cosmetic. If Eq. 12 is a typo, the actual implementation must be specified; as printed, all M.'s I numbers are uninterpretable, and the spatial-coherence half of the central claim is quantitatively unsupported. The EPE results in Table 1 are independent and may survive, but the dynamic-scene-reconstruction coherence claims need the metric fixed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a spline-based trajectory representation for dynamic point and scene reconstruction. Trajectories are modeled as cubic Hermite splines whose knot positions and tangents are predicted by a coordinate network; temporal conditioning is implemented through a low-rank time-variant spatial encoding (TVSE) applied to MLP, triplane, and triaxis backbones. Analytical derivatives of the spline yield velocity and acceleration regularizers. Experiments on DeformingThings4D scene-flow interpolation, NeRF-DS, Hyper-NeRF, Neu3D, and D-NeRF report improvements in endpoint error and a Moran's I-based spatial-coherence metric. The paper claims that explicit spline interpolation outperforms deep feature interpolation under sparse temporal sampling and improves spatial coherence without LBS or ARAP constraints.","tokens_in":23194,"tokens_out":6056,"duration_ms":64609,"significance":"If the results hold, the paper offers a simple, encoding-agnostic way to inject explicit temporal continuity into implicit deformation fields, with a clean degrees-of-freedom interpretation and cheap analytical regularizers. The reported EPE gains on DeformingThings4D (40.74/68.28 versus 45.72/73.00 for ResFields with AIAP, Table 1) are substantial and are independent of the coherence metric. The manuscript also provides detailed architecture descriptions and ablations. However, the paper's spatial-coherence evidence is compromised by the non-standard Moran's I formula in Eq. 12; until that is corrected, only the EPE-based interpolation claim is quantitatively supported.","major_comments":[{"comment":"The quantity called Moran's I in Eq. (12) is not Moran's I as defined in the cited reference (Moran 1950). The standard statistic is I = (N/W) * sum_i sum_j w_ij (z_i - zbar)(z_j - zbar) / sum_i (z_i - zbar)^2. Equation (12) instead computes K/(sum_i sum_j w_ij) * (sum_i sum_j w_ij v_i dot v_j)/(sum_i v_i), using raw 3D offsets, no mean centering, and a denominator that is neither squared nor scalar. For vector-valued offsets this is dimensionally inconsistent, and an uncentered dot-product index is dominated by any common translation, so it can report high autocorrelation even for spatially random velocities superimposed on a global offset. Because Tables 1-4 and Figures 5-7 use this index to support the spatial-coherence half of the central claim, the printed formula makes those numbers uninterpretable. Please replace Eq. (12) with a correctly centered statistic (or a clearly specified and dimensionally consistent vector generalization), recompute all Moran's I values, and state explicitly whether the implementation matches the corrected formula.","section":"A.1, Eq. (12)"},{"comment":"The paper asserts that Eq. (1) with K=2, i.e., N=T/2, guarantees a well-determined fit and removes heuristic knot selection. This is contradicted by Sec. 3.2.1, where N=30 is set manually for NeRF-DS ('To account for the motion characteristics in this dataset, we set N = 30 and rank = 10'), and by App. D, which admits that 'the dense number of knots increases the risk of jittering in certain corner cases.' If N is a per-dataset hyperparameter, the advantage of the explicit representation is partly a capacity choice rather than a principled consequence of degree-of-freedom counting. Please reconcile the two statements, justify the manual N=30 choice in terms of the effective number of training timesteps, and provide a sensitivity study over N.","section":"Sec. 3.2.1 vs. Eq. (1) and App. D"},{"comment":"Equations (6) and (8) are derivatives with respect to the normalized time \\bar{t}; the physical velocity and acceleration are (1/tau) v(\\bar{t}) and (1/tau^2) a(\\bar{t}), where tau = 1/(N-1) is the knot interval. Omitting these factors means the regularization losses L_v and L_acc have strengths that implicitly depend on N, and the claim that v has 'physical meaning' is imprecise. If the normalization is intentional, state it explicitly; if physical units are intended, include the interval factors in the loss definitions so that comparisons across different N and across datasets are not confounded by the missing 1/tau factors.","section":"Sec. 2.3, Eqs. (6) and (8)"}],"minor_comments":[{"comment":"The caption contains a typo: 'DeformaingTings4D' should be 'DeformingThings4D', and the notation 'M. ’sI' is awkwardly formatted throughout the tables.","section":"Table 1 caption"},{"comment":"The index notation in Eq. (12) is unclear: the sums run over K neighbor indices, but the quantity is written as I_t(x_i), and the relationship between the outer point i and the neighbor set should be clarified.","section":"A.1, Eq. (12)"},{"comment":"The T-Rex row contains '34.22 FALSE 0.73' in the PSNR/SSIM/Moran's I columns, which appears to be a formatting error; these columns should contain numerical values or a clear explanation for missing entries.","section":"Table 9, T-Rex row"},{"comment":"The caption says 'preserve the handle with reflection batter than the baselines'; 'batter' should be 'better'.","section":"Fig. 7 caption"},{"comment":"The text contains typos, including 'ResFileds' (should be 'ResFields') and 'concept on' (should be something like 'build on'); please proofread for these and similar errors.","section":"Sec. 1 and Sec. 3.1"},{"comment":"The hyperparameter table (Table 6) lists alpha and beta per dataset, but the text also mentions learning-rate multipliers for grids and MLPs; adding these to the table or to the text would improve reproducibility.","section":"A.3"}],"recommendation":"major_revision","confidential_remarks":"The Moran's I issue is the main risk to the paper's spatial-coherence claims; the EPE results in Table 1 are independent and appear to support the interpolation claim. If the authors can respecify the metric and recompute all coherence numbers, the paper could become publishable. The N-selection contradiction is also important but fixable with a sensitivity study. I do not see deeper circularity in the evaluation: held-out timesteps are genuinely interpolated, and the velocity/acceleration losses act as regularizers rather than as predictions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the spline deformation field is a good idea, and the EPE numbers on sparse temporal interpolation are real. But the Moran's I metric in A.1 is not Moran's I, and the spatial-coherence claims rest on it. That needs to be fixed before you can trust the coherence comparisons.\n\nThe genuinely new pieces: expressing trajectories as cubic Hermite splines whose knots and tangents are predicted by a coordinate network, then using the analytic derivative as a velocity regularizer and the second derivative as an acceleration penalty. The low-rank time-variant encoding is a sensible extension of ResFields to grids and MLPs. On DeformingThings4D with every 4th/6th frame held out, the method gets EPE 40.74/68.28 versus ResFields' best 45.72/73.00 with AIAP. That's a real improvement, and the method also looks like it could be a practical baseline. The paper is honest about its limitations on D-NeRF and shows failure cases, which I appreciate.\n\nThe soft spots. First, the metric. Equation 12 in the appendix defines I_t(x_i) with raw displacement vectors, no mean centering, and a denominator that is neither squared nor a variance. That is not Moran's I by any standard definition. An uncentered dot-product index can be dominated by global motion and will report high autocorrelation even for spatially random velocities on a common offset. Every M.'s I number in Tables 1-4 is therefore uninterpretable as printed. If the actual implementation is different, the paper must say so; as printed, the spatial-coherence half of the central claim is unsupported. The EPE results survive this, but only half the argument stands.\n\nSecond, the DoF argument for N=T/2 is undercut by the manual N=30 on NeRF-DS. That is fine in practice, but the paper does advertise removing heuristic knot selection, and it doesn't. Third, single runs and no code: not fatal, but it would help to see variance or at least a public implementation.\n\nWho this is for: anyone working on 4D reconstruction or dynamic scene flow will find the spline representation worth building on. It deserves a serious referee. A referee should require the metric to be fixed and the knot-selection story to be clarified. I'd engage with it.","headline":"Solid spline-based deformation representation with real interpolation gains, but the Moran's I metric as printed is not Moran's I and the coherence claims need a rework.","tokens_in":23702,"tokens_out":3835,"would_cite":true,"duration_ms":36144,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Representing dense point trajectories as cubic splines, with knot positions and tangents predicted by a coordinate network, beats implicit deep-feature interpolation on sparse temporal signals and improves spatial coherence in dynamic…","keywords":["spline interpolation","deformation fields","dynamic scene reconstruction","sparse temporal signals","cubic Hermite splines","low-rank time-variant encoding","spatial coherence","Moran's I"],"falsifier":"Train the same spline pipeline on DeformingThings4D with every 4th frame while sweeping N above and below T/2 (say N=T/4 and N=T); if some off-rule N matches or beats N=T/2 in EPE and Moran's I, or if the optimal N shifts with motion complexity, the degrees-of-freedom rule is empirically falsified.","tokens_in":22690,"feed_emoji":"📐","tokens_out":6599,"duration_ms":67684,"temperature":0.7,"pith_summary":"Deformation fields for dynamic 3D scenes are usually implicit neural functions that map coordinates to motion offsets; those functions are smooth by construction but can drift when input frames are sparse. This paper tries to establish that an explicit spline representation—trajectories as cubic Hermite splines whose knots and tangents are predicted by a coordinate network—interpolates sparse temporal signals better than deep-feature interpolation, and that its analytically derived velocity and acceleration make dynamic reconstruction more spatially coherent. On DeformingThings4D long sequences trained on every 4th or 6th frame, the method reports end-point error 40.74/68.28, below the best ResFields setting 45.72/73.00 with AIAP regularization. If correct, this means explicit trajectory geometry and a decoupled low-rank time-variant encoding can replace heuristics like linear blend skinning or as-rigid-as-possible constraints in ill-posed dynamic reconstruction.","feed_headline":"Explicit splines beat implicit neural interpolation on sparse motion","feed_subtitle":"Cubic-spline trajectories cut error on sparse-frame interpolation and yield cleaner 3D scene reconstruction.","key_machinery":"The load-bearing mechanism is the spline deformation field: a cubic Hermite spline between a start and end knot, where each knot carries a position and a tangent predicted by a coordinate network. The network uses time-variant spatial encoding: at timestep t, spatial features are decoded through base weights plus a low-rank residual sum over temporal codes, decoupling spatial and temporal information. Because the spline is polynomial in normalized time, velocity and acceleration are obtained by differentiating the interpolation function; these analytic derivatives feed a neighbor-consistent velocity loss and an acceleration magnitude loss. The knot count is tied to the training timestep count by K·N=T with K=2 for cubic Hermite, so N=T/2, which the paper argues makes the fit well-determined. This machinery turns trajectory modeling into an interpolation problem with explicit, physically meaningful derivatives rather than implicit network smoothness.","core_discovery":"The paper's central object is a cubic Hermite spline applied to each point's trajectory over time: at knots t_start and t_end the network predicts both position offsets and tangents; query times are interpolated by a third-order polynomial. Choosing N=T/2 knots for T training timesteps is proposed as a degrees-of-freedom balance that avoids under- and over-fitting; the closed-form derivative of the spline then supplies velocity, used in a neighbor-consistency loss L_v, and acceleration, used as a jitter suppression loss L_acc. A low-rank time-variant spatial encoding (TVSE) replaces coupled 4D inputs, letting the same encoding run with sinusoidal positional encodings, dense grids, or residual field weights. The resulting spline field is evaluated on scene-flow interpolation and dynamic scene reconstruction with Gaussian splatting, where it outperforms deep-feature interpolation on sparse temporal inputs and raises a Moran's I spatial-coherence measure across NeRF-DS, Hyper-NeRF, and Neu3D.","pith_inferences":["If the N=T/2 rule is genuine capacity accounting, knot count should generalize across sequences of different lengths; the paper's manual N=30 for NeRF-DS suggests the rule may instead be a tuned prior, and a knot sweep across datasets would clarify the distinction.","Because TVSE is architecture-agnostic, the spline representation could be combined with hash grids or future encodings; the supplementary material already suggests time-variant spatial hash encoding as promising.","Analytic velocity opens a natural interface to physics-based losses or simulation, a direction the paper only gestures at through advection experiments.","Moran's I, as defined here, measures spatial autocorrelation of discrete motion vectors; treating it as a standard coherence metric for dynamic scene reconstruction would require checking its agreement with perceptual quality on more diverse scenes."],"forward_implications":["On long-sequence scene-flow interpolation with every 4th or 6th frame as input, the spline field reports EPE 40.74/68.28, beating ResFields with AIAP at 45.72/73.00 and DOMA at 138.80/146.51.","The explicit regularity terms extend to both MLP and grid backbones, so SF-PE-ResFields, SF-Triplanes, and SF-Triaxes all raise Moran's I on NeRF-DS, Hyper-NeRF, and Neu3D without linear blend skinning or as-rigid-as-possible constraints.","Analytic velocity enables motion advection and trajectory editing with only keyframe edits, and coherent canonical-space deformations allow style transfer across the full sequence.","The knot-counting rule N=T/2 is presented as eliminating heuristic knot selection and as a diagnostic tool: reducing N in large-motion data produces large misalignment that penalizes reconstruction accuracy."],"supporting_citations":[{"why":"Baseline and design source: ResFields injects temporal information into network weights, which TVSE generalizes; its reported EPE numbers are the comparison to beat.","marker":"[Mihajlovic et al. 2024]"},{"why":"Provides the DOMA scene-flow interpolation protocol, sparse-supervision setting, and SIREN backbone used in the interpolation experiments.","marker":"[Zhang et al. 2024]"},{"why":"4DGS is the dynamic Gaussian splatting pipeline that grid-based variants extend and a principal reconstruction baseline.","marker":"[Wu et al. 2023]"},{"why":"Supplies the plane/axis-based grid encoding and element-wise feature aggregation used by SF-Triplanes and SF-Triaxes.","marker":"[Fridovich-Keil et al. 2023]"},{"why":"HexPlane is the grid representation the time-variant triplane variants build on.","marker":"[Cao and Johnson 2023]"},{"why":"DeformingThings4D is the dataset for long-sequence scene-flow interpolation evaluation.","marker":"[Li et al. 2021]"},{"why":"Defines Moran's I, which the paper adapts into its spatial-coherence metric.","marker":"[Moran 1950]"},{"why":"Spline-NeRF is cited as prior inspiration for representing trajectories with splines.","marker":"[Knodt 2022]"},{"why":"Neural spline fields are cited as prior work motivating spline trajectory representations.","marker":"[Chugunov et al. 2023]"}],"fun_headline_variants":["Spline knots beat neural fields on sparse motion","Cubic-spline trajectories sharpen sparse interpolation","Explicit splines improve spatial coherence in dynamic scenes","Low-rank time-variant encoding boosts spline fields","Spline deformation: fewer knots, cleaner motion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that tying the knot count to N=T/2 makes the spline fit well-determined without heuristic tuning, yet for NeRF-DS the paper sets N=30 by hand; if the right N is dataset-dependent, the claimed advantage of explicit interpolation is partly a capacity choice.","fun_headline_variants_meta":{"raw":{"variants":["Spline knots beat neural fields on sparse motion","Cubic-spline trajectories sharpen sparse interpolation","Explicit splines improve spatial coherence in dynamic scenes","Low-rank time-variant encoding boosts spline fields","Spline deformation: fewer knots, cleaner motion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000662,"raw_usage":{"total_tokens":3027,"prompt_tokens":947,"completion_tokens":2080,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":2008}},"tokens_in":563,"tokens_out":2080,"duration_ms":16951,"temperature":1.0,"reasoning_tokens":2008,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T18:38:33.891758+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same spline pipeline on DeformingThings4D with every 4th frame while sweeping N above and below T/2 (say N=T/4 and N=T); if some off-rule N matches or beats N=T/2 in EPE and Moran's I, or if the optimal N shifts with motion complexity, the degrees-of-freedom rule is empirically falsified.","supporting_citations":[],"review_version":1}