{"id":"4be22891-7193-40f5-bdba-65164f9765fa","arxiv_id":"1909.00767","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"A NURBS-surface model with scaling, and optionally weight, estimation tracks a vehicle's 3D extent and kinematics in real time from sparse LiDAR points, outperforming a 3D Gaussian-process tracker on shape area error.","lead":"This paper presents two LiDAR-based methods that track a vehicle's 3D shape and motion simultaneously using flexible NURBS surfaces. Both methods run fast enough for real-time use and, in real-world tests, estimate shape more accurately than a state-of-the-art 3D Gaussian-process tracker.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The inside/outside test for the NURBS shape function is circular and unstated: Eq.22 defines d through membership in S_d, while Eq.5 defines S_d through d≥0; every UKF pseudo-measurement inherits this undefined sign.","rationale":"The paper's central claim—joint 3D shape and dynamics estimation with real-time NURBS tracking—stands or falls on the measurement model. The reader's weakest assumption identifies the circular definition of the inside set S_d in Eqs.5 and 22. I agree that this is the most load-bearing weakness: the UKF's stacked pseudo-measurement (Eqs.6–9) requires the signed shape function at every measurement, and no algorithm is given for the required point-in-surface test. Without that test, the sign of d, and hence the entire observation update, is under-specified. I also note a second, related issue that strengthens the concern: Eq.21 defines d as a Mahalanobis distance to the closest surface point, but the level-set RHM in Eq.6 uses α as a scaling factor; for non-circular shapes these are different models. The evaluation does not close the gap because it reports only 2D area error and a small set of scalar RMSEs, not a volumetric or surface error. These are correctness and reproducibility risks, not evidence of fraud. The methods are plausible, the real-world comparisons are useful, and the missing pieces are likely implementable in a straightforward way, so rejection is not warranted. The appropriate disposition remains CONDITIONAL: the manuscript should be accepted only if the sign test and the scaling interpretation of d are specified, and ideally if the implementation is released. My read therefore leaves the reader's CONDITIONAL verdict unchanged.","tokens_in":9039,"tokens_out":12141,"duration_ms":140930,"concrete_test":"Re-run the static scenario (Section V-A) with M2, adding one explicit inside/outside rule to Eq.22: for a measurement in target coordinates, compute the intersection of the ray from m_k through z with the NURBS surface; if the surface intersection is farther from m_k than z, declare z inside (d = +m), otherwise outside (d = −m). Compare the resulting Table I values with the published EA=1.433 and Epos=0.393. In parallel, re-derive the implied source distribution for a unit sphere: with R=I and dmax=1, Eq.6 combined with Eq.21 yields source radius 1−α, not the radial scaling α of a level-set RHM; if the authors intended radial scaling, Eq.21 must be replaced by a radial-ratio shape function. A change in EA beyond a small tolerance, or a sphere test showing the mismatch, confirms the concern.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The measurement model in Eq.6 requires evaluating d(x_k, z_kl) for each measurement. Eq.22 sets d = +m if z∈S_d and d = −m otherwise, with S_d defined in Eq.5 as {z | d(x,z) ≥ 0}. This is circular: the sign of d is defined by membership in a set that is itself defined by the sign of d. The manuscript never specifies an independent geometric point-in-surface test, such as ray casting, winding number, or radial comparison against the NURBS surface. This is not a cosmetic gap because Eq.6 is the only observation model feeding the UKF; if the sign convention is wrong, the pseudo-measurement 0 = α·dmax − d is shifted by 2m, biasing the estimated position, orientation, and surface. A related inconsistency: Eq.6 is taken from level-set RHM, where α acts as a radial scaling of the shape, but Eq.21–22 define d as a closest-surface Mahalanobis distance whose level sets are inward offsets, not scaled copies of the boundary. For a box-like vehicle these two parameterizations disagree, so the generative model behind Eq.6 should be justified. The paper's own limitations section (V-C) discusses process-noise initialization but never addresses this sign/inside ambiguity, and Table I's area error does not isolate 3D shape bias.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two NURBS-surface-based methods for 3D extended target tracking from LiDAR point clouds. Method M1 estimates per-axis scaling factors and NURBS weights, while Method M2 estimates scaling factors only. The target state includes kinematics (position, orientation, velocity, curvature) and shape parameters, and a UKF is used for inference. The measurement model is based on a level-set random hypersurface model, with a signed Mahalanobis distance to the closest surface point as the shape function. The methods are evaluated on real-world static and dynamic vehicle scenarios against 3D Gaussian process, 2D B-spline, and point tracking baselines, reporting lower area error and runtimes of 3.0-18.8 ms per step.","tokens_in":9364,"tokens_out":6939,"duration_ms":65377,"significance":"If the results hold, the paper offers a real-time capable 3D extended target tracking method that jointly estimates shape and kinematics, demonstrated on real LiDAR data. The evaluation is thorough in terms of real-world data and comparison with state-of-the-art methods. The paper's explicit limitation discussion and reproducible parameter settings are strengths. However, the measurement model and shape function definition contain circular and underspecified steps that currently prevent full verification.","major_comments":[{"comment":"Eqs. (5) and (22) define S_d(x_k) and the sign of d(x_k,z) circularly: S_d is the set with d >= 0, and d is positive exactly for points in S_d. Moreover, dmax in Eq. (6) is the maximum of d over this set, so it inherits the same ambiguity. The manuscript never specifies an independent geometric test for determining whether a measurement lies inside the NURBS surface (e.g., ray casting, winding number, or a signed projection onto the surface normal). Because Eq. (6) is the only observation model feeding the UKF, an unstated or inconsistent test will bias the pseudo-measurement by 2m and can collapse the shape estimate. Please provide the concrete test used and define dmax accordingly.","section":"Eqs. (5) and (22)"},{"comment":"The level-set RHM scaling relation in Eq. (6) assumes that the level sets of the shape function are scaled copies of the boundary. The shape function defined in Eqs. (21)-(22) as the signed closest-surface Mahalanobis distance has level sets that are inward/outward offsets, which are not scaled copies for a general NURBS surface (e.g., a box-like vehicle). The paper does not justify using the offset distance inside the scaled-copy measurement model; this is a fundamental model mismatch that should be either justified or corrected by redefining the shape function (for example, as a radial function in the star-convex framework).","section":"Eqs. (6), (21)-(22)"},{"comment":"The weight dynamics in Eq. (18) do not specify which surface parameter point (u,v) is associated with the weight omega_{ij}. The paper should state the mapping from control-point indices to parameter values (e.g., Greville abscissae or knot averages) and define the Gaussian curvature argument accordingly. Without this, the evolution of the weights is ambiguous and the regularization process is not reproducible.","section":"Eq. (18)"},{"comment":"The limitations discussion does not address the circular inside/outside definition or the offset-versus-scaled level-set issue. These are not cosmetic gaps; they directly affect the measurement model and the meaning of the pseudo-measurement. The manuscript should include an explicit statement of the geometric point-in-surface test used and the assumptions under which Eq. (6) is an appropriate measurement generator.","section":"Section V-C"}],"minor_comments":[{"comment":"In Eq. (9), the stacked vector elements are scalars, so the transpose on each g(...) is unnecessary and notationally confusing. Please stack the scalars directly as a column vector.","section":"Eq. (9)"},{"comment":"The optimization in Eq. (20) is not described; please give the algorithm (e.g., Newton-Raphson or a coarse-to-fine search) and stopping criteria, as this step dominates the runtime and influences the registration accuracy.","section":"Eq. (20)"},{"comment":"The area error metric is defined as the area of the encasing rectangle of the estimated shape; this is a 2D measure and does not capture 3D shape accuracy. Please clarify whether this is intentional and discuss its implications for assessing the 3D shape estimation.","section":"Tables I and III"},{"comment":"Please specify how the random sampling of the 50 (static) and 20 (dynamic) points is performed (e.g., uniform over the point cloud) and how the convex-hull points are chosen, since the number of points affects the reported runtimes and errors.","section":"Sections V-A and V-B"},{"comment":"The limitations paragraph mentions that M2 is limited to approximately cuboid objects, but it would be useful to also state the constraints of M1 with respect to shape complexity and the chosen NURBS parameterization.","section":"Section V-C"},{"comment":"In Figure 6 the legend includes GT, but Figure 5 does not; please add the GT curve to the static-scenario plots or explain its absence.","section":"Figures 5 and 6"},{"comment":"The paper cites reference [10] for level-set RHM; consider also citing the original star-convex RHM work [9] in the derivation of Eq. (6) to clarify the intended level-set scaling interpretation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal scope. The real-world evaluation is a strength, but the measurement model issues are serious and need to be addressed before publication. The authors should be asked to supply the missing geometric test and to clarify/justify the level-set scaling model."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: this is a legitimate incremental contribution — the first to put NURBS surfaces into level-set RHM for 3D extended target tracking, with scaling-only and scaling-plus-weights variants, and a real-world LiDAR evaluation showing real-time runtimes and lower area error than a 3DGP baseline. Worth reading for anyone doing automotive ETT.\n\nWhat's good: the NURBS surface representation is natural for vehicles; the two methods give a practical speed/accuracy trade-off (M2 at 3 ms vs 3DGP at 118 ms with 50 points). They honestly report scenarios with occlusions and segmentation errors, and the limitations section is candid about process-noise sensitivity. The comparison against [20] is fair — they correctly note that the 2D B-spline method cannot handle 3D data.\n\nSoft spots, in order of seriousness:\n\nThe measurement model is under-specified. Eq. 22 defines the signed distance d through membership in the inside set S_d, and Eq. 5 defines S_d as {z | d ≥ 0}. That is circular on the page. Any implementation has to break the circle with an actual point-in-surface test (ray casting, radial comparison, etc.), but the paper never says which one. This is not just cosmetics: the UKF pseudo-measurement in Eq. 6 depends on the sign of d, and getting it wrong flips the update. A short paragraph specifying the geometric test would fix it.\n\nSecond, the weight dynamics in Eq. 18 do not state which surface parameter (u,v) corresponds to weight ω_ij. A control point's weight is associated with a basis function, not a single surface point; the regularization formula needs a defined mapping. Also missing: initial control point coordinates. The figures show the initial surfaces, but the actual coordinates are absent, so the experiments cannot be reproduced from the text.\n\nThird, the results have no error bars or multiple runs, so the area-error gaps (e.g., EA 0.617 vs 5.789 in static) are single-shot numbers. The runtime table is useful but hardware and implementation details are thin.\n\nThe stress-test note also flags a conceptual inconsistency between the level-set RHM scaling (α · dmax, which scales the shape radially) and the Mahalanobis distance in Eq. 21, whose level sets are inward offsets rather than scaled copies. For boxy vehicles these two parameterizations can disagree. This is a real issue, but I think it is secondary to the missing inside/outside test; a careful derivation or a clarifying paragraph would address it.\n\nBottom line: the core idea is sound and the experiments are relevant, but the model description has reproducibility gaps. A serious referee should engage; the paper deserves a major-revision request, not a desk reject. This is a solid 'conditional accept' as is.","headline":"Legitimate 3D extended target tracking extension using NURBS surfaces with real-world evaluation, but the measurement model is under-specified and the write-up needs a major revision for reproducibility.","tokens_in":9953,"tokens_out":2236,"would_cite":true,"duration_ms":23002,"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":"This paper claims that a vehicle's full 3D shape and dynamics can be estimated together in real time from sparse LiDAR points by modeling the target as a NURBS surface whose scaling parameters and optionally weights are part of the filter…","keywords":["extended target tracking","NURBS surfaces","LiDAR point clouds","unscented Kalman filter","3D shape estimation","random hypersurface model","vehicle tracking","real-time estimation"],"falsifier":"A controlled synthetic test with known ground truth: generate LiDAR points just inside and just outside a known NURBS surface whose true surface and labels are known, run both proposed filters, and check whether a single marginal point whose inside/outside label is deliberately flipped causes the estimated surface to drift or the area error to jump; if the hidden point-in-surface test is inconsistent, the sign of $d$ flips and the pseudo-measurement is biased.","tokens_in":1796,"feed_emoji":"🚗","tokens_out":7435,"duration_ms":107315,"temperature":0.7,"pith_summary":"This paper sets out to show that the full three-dimensional shape of a moving vehicle can be tracked jointly with its position, velocity, and orientation from sparse, noisy LiDAR measurements. The shape is represented as a NURBS surface, a standard CAD surface model, whose axes are scaled, and optionally whose control-point weights are adjusted, by an unscented Kalman filter. Two methods are proposed: one estimates both scaling and weights for better shape fidelity, and a faster one estimates only scaling. On real sensor data of a static and a moving vehicle, both methods match or improve on a state-of-the-art 3D Gaussian-process tracker while running in 3.0-18.8 ms per time step. If the claim holds, autonomous vehicles can use a compact parametric model of an entire 3D obstacle rather than a 2D footprint or bounding box.","feed_headline":"NURBS surface tracks a car's full 3D shape from LiDAR in 3.1 ms","feed_subtitle":"Two new filters estimate extent and motion together, beating 3D Gaussian-process tracking on real vehicle data","key_machinery":"The load-bearing object is the NURBS surface $S_k(u,v,\\omega_k,s_k)$ of Eq. (12): a tensor-product rational B-spline whose three Cartesian scaling factors $s_k=(s_x,s_y,s_z)$ and control-point weights $\\omega_k$ are part of the tracking state. Around it, the measurement model uses the level-set random hypersurface equation $0 = \\alpha_{kl}\\, d_{\\max}(x_k) - d(x_k, y_{kl}-w_{kl})$ (Eq. 6), where $d$ is a signed Mahalanobis distance to the closest surface point found by minimizing angle difference (Eqs. 20-22). The sign of $d$ is set by membership in the inside set $S_d(x_k)$, so the entire filter update reduces to repeatedly evaluating whether a point lies inside or outside the current NURBS surface. Inference is done with an unscented Kalman filter that processes all measurements in one stacked update to keep the posterior order-independent.","core_discovery":"The central claim is that a NURBS surface can serve as the shape model in an extended target tracking filter, so that the target's unknown extent and its kinematics are recovered from the same noisy point measurements. The measurements enter through a level-set random hypersurface measurement equation, where each LiDAR point is associated with the surface point closest in angle, and the residual is a signed Mahalanobis distance to that surface point. Method 1 augments the kinematic state with three Cartesian scaling factors plus all NURBS weights, regularized by normalized Gaussian curvature; Method 2 estimates only the three scaling factors. In the reported real-world scenarios, Method 1 reaches an area error of $0.617\\,\\text{m}^2$ (static) and Method 2 reaches $0.323\\,\\text{m}^2$ (dynamic), compared with $5.789\\,\\text{m}^2$ and $3.576\\,\\text{m}^2$ for the 3D Gaussian-process baseline, with mean runtimes of $18.8\\,\\text{ms}$ and $3.1\\,\\text{ms}$ per 50-point update. The authors conclude that both methods are real-time capable and estimate a fully 3D extent rather than a 2D approximation.","pith_inferences":["The same NURBS machinery could be extended to non-convex targets by replacing the angle-difference closest-point criterion of Eq. (20) with a minimum-distance criterion; the paper notes this possibility in passing but does not implement it.","A practical implementation needs a concrete inside/outside test, such as ray casting to the surface or checking convex hull membership, and the choice of that test will likely dominate tracking quality near the surface.","The fixed process-noise trade-off described in the limitations suggests that an adaptive process-noise schedule for the shape states would be a natural next step, especially for large errors in the initial velocity estimate."],"forward_implications":["A vehicle's entire 3D extent, not just a 2D footprint, can be estimated in real time from a roof-mounted LiDAR, with each 50-point update taking 18.8 ms for Method 1 and 3.1 ms for Method 2.","The method continues to estimate a useful shape under severe occlusion, since the dynamic scenario's measurements came only from the back of the car.","Estimating NURBS weights in addition to scaling improves static-shape accuracy (area error 0.617 vs 1.433) at the cost of about six times the runtime.","The scaling-only method is claimed to fit approximately cuboid objects such as vehicles, pedestrians, and bicycles.","Because all measurements are incorporated in one stacked UKF update, the resulting estimate is order-independent with respect to the order of the measurements."],"supporting_citations":[{"why":"Supplies the level-set random hypersurface measurement model that connects the shape function to the noisy measurements.","marker":"[10]"},{"why":"Provides the NURBS surface definition and tensor-product basis used as the shape representation.","marker":"[3]"},{"why":"The 3D Gaussian-process extended object tracker used as the main state-of-the-art baseline in the evaluations.","marker":"[13]"},{"why":"The closest prior spline-based vehicle model, which the paper extends from 2D B-spline curves to 3D NURBS surfaces with dynamics.","marker":"[20]"},{"why":"The unscented Kalman filter used for inference with the nonlinear measurement model.","marker":"[23]"},{"why":"The constant curvature and velocity motion model used for the kinematic part of the state.","marker":"[21]"},{"why":"Provides the Gaussian curvature regularization approach that smooths the estimated NURBS weights in Method 1.","marker":"[24]"}],"fun_headline_variants":["NURBS filters track 3D car shape from LiDAR in 3.1 ms","Fast 3D target tracking: NURBS beats GP in speed and accuracy","LiDAR target tracking with NURBS: 6x faster than GP baseline","Full 3D extent from LiDAR: NURBS tracking at 3.1 ms per frame","Joint shape and motion from sparse LiDAR using NURBS surfaces"],"cache_read_input_tokens":11904,"weakest_assumption_plain":"The filter assumes that each measurement can be reliably classified as lying inside or outside the currently estimated surface, because that classification sets the sign of the distance residual; the paper never specifies how this classification is done.","fun_headline_variants_meta":{"raw":{"variants":["NURBS filters track 3D car shape from LiDAR in 3.1 ms","Fast 3D target tracking: NURBS beats GP in speed and accuracy","LiDAR target tracking with NURBS: 6x faster than GP baseline","Full 3D extent from LiDAR: NURBS tracking at 3.1 ms per frame","Joint shape and motion from sparse LiDAR using NURBS surfaces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000249,"raw_usage":{"total_tokens":1535,"prompt_tokens":918,"completion_tokens":617,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":506}},"tokens_in":534,"tokens_out":617,"duration_ms":264616,"temperature":1.0,"reasoning_tokens":506,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:37:28.614597+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled synthetic test with known ground truth: generate LiDAR points just inside and just outside a known NURBS surface whose true surface and labels are known, run both proposed filters, and check whether a single marginal point whose inside/outside label is deliberately flipped causes the estimated surface to drift or the area error to jump; if the hidden point-in-surface test is inconsistent, the sign of $d$ flips and the pseudo-measurement is biased.","supporting_citations":[{"cited_title":"Comparison and Evaluation of Advanced Motion Models for Vehicle Tracking,","cited_arxiv_id":null,"evidence_quote":"The constant curvature and velocity motion model used for the kinematic part of the state."},{"cited_title":"Level-Set Random Hypersurface Models for Tracking Non-Convex Extended Objects,","cited_arxiv_id":null,"evidence_quote":"Supplies the level-set random hypersurface measurement model that connects the shape function to the noisy measurements."},{"cited_title":"On NURBS: A Survey,","cited_arxiv_id":null,"evidence_quote":"Provides the NURBS surface definition and tensor-product basis used as the shape representation."},{"cited_title":"3D Extended Object Tracking Using Recursive Gaussian Processes,","cited_arxiv_id":null,"evidence_quote":"The 3D Gaussian-process extended object tracker used as the main state-of-the-art baseline in the evaluations."},{"cited_title":"A Cartesian B-Spline Vehicle Model for Extended Object Tracking,","cited_arxiv_id":null,"evidence_quote":"The closest prior spline-based vehicle model, which the paper extends from 2D B-spline curves to 3D NURBS surfaces with dynamics."},{"cited_title":"The Unscented Kalman Filter for Nonlinear Estimation,","cited_arxiv_id":null,"evidence_quote":"The unscented Kalman filter used for inference with the nonlinear measurement model."},{"cited_title":"Reconstruction of free-form space curves using NURBS-snakes and a quadratic programming approach,","cited_arxiv_id":null,"evidence_quote":"Provides the Gaussian curvature regularization approach that smooths the estimated NURBS weights in Method 1."}],"review_version":1}