{"id":"8f4370f6-c187-4b92-b737-02a881a4fd05","arxiv_id":"2505.05132","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"An active contour model jointly optimizes cubic Bézier curve endpoints, tangent orientations, and control points, reducing vectorization error on silhouettes traced by Inkscape, Illustrator, and a curvature baseline.","lead":"This paper presents a post-processing method that improves how computer programs convert the outline of a bitmap image into smooth curved vector graphics. It moves the end points, tangent directions, and curve shapes of Bézier curves together to hug the original boundary more tightly.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is not yet established: the non-convex alternating minimization has no convergence or robustness analysis, and the evidence is only three silhouettes with no error bars and a curvature baseline (MaxDist=6) chosen to be coarse.","rationale":"The reader's verdict is CONDITIONAL, and I agree. The reader's weakest assumption was convergence of the non-convex alternating minimization to a useful local minimum; my concern is that, plus the generalizability of the very small evaluation. The paper's mathematical formulation is coherent: the objective is a geodesic active-contour energy, the Bezier parameterization with tangent constraints is clearly described, and the reported reductions are consistent across three initializers. I do not see an internal inconsistency in the derivation. However, the empirical claim is under-supported: no code or data release, no error bars, a baseline parameter (MaxDist=6) explicitly chosen to make the initial curvature-method fit coarse, and no analysis of sensitivity to the local search windows (t ± 2, α ± 4 degrees) or to the stopping rule. A single concrete robustness and generalization check would settle whether the central claim holds broadly. Therefore the appropriate verdict remains CONDITIONAL; my read does not change the reader's verdict.","tokens_in":12830,"tokens_out":18694,"duration_ms":200915,"concrete_test":"Run the full pipeline on 100 public-domain binary silhouettes (e.g., a random subset of MPEG-7 CE-Shape-1) using the curvature-method and Inkscape initializers with the paper's parameters, and record for each shape the relative changes in d(B,C), d(C,B), and total Bezier length before and after. Report the mean and 95% bootstrap confidence interval of the relative changes and the fraction of shapes where d(C,B) worsens. If the confidence interval includes zero or a non-negligible fraction of shapes worsen, the claim that the active contour 'significantly reduces' the distance to the silhouette boundary for any vectorization is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the alternating minimization of the non-convex energy (6), as implemented in Section 4, reaches a useful local minimum from each initial vectorization, and that the three test silhouettes are representative. This premise is insecure for two reasons. First, the optimization is heuristic: each t_n is updated only within [t_n−2, t_n+2] and each α_n within ±4 degrees, the inner Bezier-parameter problem (7) is solved by gradient descent from a linear estimate (8), and iteration stops when the energy is no longer 'significantly reduced.' No convergence theorem, no study of sensitivity to the window sizes, and no random-restart or robustness experiments are given. Second, the empirical support is thin: three silhouettes, no error bars or significance tests, and the curvature-method baseline uses MaxDist=6, which the Appendix states was chosen to give 'a few interpolation values allowing a better illustration of the performance.' Consequently, the claimed 15–54% reductions may reflect favorable initializers and parameter choices rather than a general property of the active contour model. The paper is honest about the asymmetric metrics and reports d(C,B) as well, but those d(C,B) improvements also come from the same three cases only.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a variational active contour model that refits a collection of cubic Bézier curves to a closed silhouette boundary C(t). The energy in Eq. (6) is a geodesic-active-contour integral of the distance to C, weighted by curve length; the unknowns are the curve endpoints t_n, tangent orientations α_n at regular points, and the per-interval Bézier parameters λ_n, γ_n, β_n, δ_n. Minimization is performed by alternating local updates: for fixed endpoints and orientations, the Bézier parameters are obtained by gradient descent on the per-interval functional (7), initialized by the linear least-squares estimate (8); for each n, t_n and α_n are then searched over small windows. Experiments on three silhouettes report reductions of both d(B,C) and d(C,B) for Inkscape, Adobe Illustrator, and a curvature-based baseline, and the paper demonstrates an optional length-regularization parameter w_n.","tokens_in":13102,"tokens_out":3960,"duration_ms":41659,"significance":"The active-contour formulation is clean and well motivated: the tangent-vector parameterization in Eq. (4) makes regularity constraints simple, and the per-interval decoupling for fixed (t_n, α_n) is a practical algorithmic contribution. The paper is also honest about the metric asymmetry, reporting d(C,B), which is not the minimized quantity, and explicitly discussing the symmetric alternative (12). If the reported improvements were confirmed on a broader benchmark with neutral baselines and a robustness analysis, the method would be a useful post-processing tool for silhouette tracing. The main weakness is that the central claim of significant improvement currently rests on three example images, a single run per method, and a heuristic optimization whose convergence is not analyzed.","major_comments":[{"comment":"The alternating minimization is heuristic: each t_n is updated only within [t_n−2, t_n+2], each α_n within ±4 degrees, the inner Bézier-parameter problem (7) is solved by gradient descent from the linear estimate (8), and the outer iteration stops when the energy is no longer 'significantly reduced.' No convergence argument, monotonicity guarantee, or sensitivity study for the window sizes is provided. Since energy (6) is non-convex, the claimed 15–54% improvements are asserted to hold from arbitrary initial vectorizations without evidence that the coordinate-descent scheme reaches a useful local minimum beyond the three reported runs.","section":"§4, Eqs. (7)–(9)"},{"comment":"The experimental support is thin. Only three silhouettes are tested, each with one run per method; no error bars, no multiple initializations, and no significance tests are reported. With nine numerical comparisons (three methods × three silhouettes), the variability is unknown, so the claim of a 'significant reduction' is not statistically established. Additional experiments on more shapes, with several restarts or perturbations, would be needed to support the generality of the improvement.","section":"§5, Table 1"},{"comment":"The curvature baseline appears calibrated to be coarse: the Appendix states that MaxDist = 6 'provides a few interpolation values allowing a better illustration of the performance.' This makes the comparison with the proposed method favorable by construction. A fair evaluation should sweep MaxDist or otherwise select the baseline at a comparable accuracy/complexity trade-off, and should report results at matched node counts or under a common complexity budget, since the number of nodes differs across methods and the paper explicitly declines to consider it.","section":"Appendix, MaxDist parameter; Table 1"},{"comment":"The primary metric d(B,C) in Eq. (10) has the same numerator as the energy (6) being minimized, so a reduction in d(B,C) is partly built into the optimization. The paper correctly acknowledges this and reports d(C,B), which is not minimized; those d(C,B) improvements provide independent evidence in principle, but they come from the same three cases and the same single runs, so the concern is only partially resolved. A larger experiment set or a comparison on the symmetric energy (12) would strengthen the case.","section":"Eq. (6), Eq. (10), Table 1"}],"minor_comments":[{"comment":"The display of the energy (6) is corrupted: the summation appears as 'NX n=1' with a broken N; please fix the typesetting.","section":"§3, Eq. (6)"},{"comment":"The column header 'Var. Perc.' should be defined explicitly (presumably 'relative variation in percent'), and the units of the distance values should be stated (likely pixels, given the 1024×1024 images).","section":"§5, Table 1"},{"comment":"The statement that (t_min, α_min) is the argmin over [t_n−2, t_n+2] × [α_n−r_α, α_n+r_α] is ambiguous because the per-interval parameters are re-optimized for each candidate pair; please clarify whether the search is performed on a discretized grid or on the continuous range using a generic optimizer.","section":"§4, Eq. (9)"},{"comment":"The stopping criterion 'significantly reduced' should be made quantitative, for example by specifying a relative energy decrease threshold or a maximum number of iterations.","section":"§4, step 2"},{"comment":"References [45] and [63] cite the same article (Caselles, Kimmel, and Sapiro, 'Geodesic Active Contours') with different identifiers; please merge them.","section":"References"},{"comment":"No code or data repository is provided, which limits reproducibility of the quantitative claims; making the three silhouette images and the parameter settings publicly available would be helpful.","section":"General"},{"comment":"The regularization effect in Fig. 5 is shown only for one zoomed section; a quantitative comparison of distances before and after applying w_n > 0 would be more informative.","section":"Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core variational formulation is sound, but the central empirical claim is not yet established because of the small benchmark, the favorable setting of the baseline parameter, and the lack of any convergence or robustness analysis for the heuristic optimizer. I would encourage the authors to expand the experiments, add sensitivity studies, and either provide a careful baseline comparison or temper the claimed generality. These are fixable within the manuscript's scope, so I do not recommend rejection at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a real step forward for one narrow problem: improving an existing Bezier trace of a binary silhouette. The authors formulate a geodesic active contour energy over a collection of cubic Bezier curves and minimize it jointly over endpoint locations, tangent orientations at regular points, and control point parameters. That combination is new as far as I can tell from the cited literature, and the endpoint-tangent parameterization is clean enough to make the regular-point constraints feel natural. That is the best part of the paper.\n\nThe experiments are the weak part. All reported improvements come from three silhouettes and three initializers, giving nine numbers. No error bars, no significance tests, no sensitivity analysis for the method's own windows. The curvature baseline's MaxDist is fixed at 6, and the appendix says this was chosen to give 'a few interpolation values allowing a better illustration' — that is a red flag for the comparison, though the authors are honest enough to print it. The primary metric d(B,C) is essentially the numerator of the energy being minimized, so its improvement is partly by construction. To their credit, they also report d(C,B), which is not minimized, and it improves in all nine cases. That is the strongest evidence the method works, but the sample is still tiny. No code or data is released, so reproducing the nine numbers means manually rerunning Inkscape and Illustrator.\n\nThe optimization is a heuristic alternating scheme with local search windows and no convergence guarantee. I would not call that fatal — this is a graphics paper, not numerical analysis — but I would want to see how the results depend on the window sizes and at least a rough runtime figure.\n\nThe central idea is sound. The overreach is in the wording: 'significantly reduces' and 'any vectorization method' go beyond what three examples support. A defensible version of this paper exists with a larger and more varied dataset, a baseline that is clearly comparable, and released code.\n\nFor a reader in vectorization or interactive graphics tools, this is a useful post-processing routine and a clean formulation. The paper deserves a serious referee, but the current evidence base should not be accepted as is.","headline":"Genuinely new Bezier active-contour formulation with an honest but much too thin experimental section; worth refereeing, not worth accepting as is.","tokens_in":13613,"tokens_out":2256,"would_cite":true,"duration_ms":24461,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D17","68U10","68U05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Minimizing a geodesic active-contour energy over cubic Bézier curves reduces silhouette vectorization error by 15 to 54 percent relative to Inkscape, Adobe Illustrator, and a curvature-based baseline on three test silhouettes.","keywords":["silhouette vectorization","cubic Bézier curves","active contour model","geodesic active contour","image tracing","curve fitting","variational methods","vector graphics"],"falsifier":"Run the active contour model on a diverse set of silhouettes using deliberately poor initial guesses and measure both $d(B,C)$ and $d(C,B)$; if for some reasonable initial guess the minimization stalls in a local minimum that leaves large uncovered boundary stretches, or if the average distance increases relative to the initial vectorization, the claim that the method significantly improves any vectorization would fail.","tokens_in":12605,"feed_emoji":"✒️","tokens_out":5412,"duration_ms":51024,"temperature":0.7,"pith_summary":"The paper proposes a post-processing step for silhouette vectorization: take any initial set of cubic Bézier curves approximating a shape boundary, and refine it by minimizing a geodesic active-contour energy that measures how far each point of the Bézier curves lies from the true silhouette. The minimization moves the curves' end points along the boundary, rotates the tangent directions at regular points, and re-estimates the Bézier control parameters. On three test silhouettes the refinement lowers the average distance between the boundary and the vectorization produced by Inkscape, Adobe Illustrator, and a curvature-based baseline by 15 to 54 percent. The interest is that the model is initialization-agnostic: any vectorization method can feed it, and its output should be closer to the original shape.","feed_headline":"Active contour pass cuts vectorization error by up to 54%","feed_subtitle":"Refines Bézier traces from Inkscape, Illustrator, and a curvature baseline on test silhouettes.","key_machinery":"The load-bearing mechanism is the geodesic active contour energy (6), a weighted curve length in which the weight at each point is its Euclidean distance to the silhouette boundary plus an optional regularization constant $w_n$. The curves are cubic Bézier segments parameterized so that each endpoint's tangent is expressed in the local frame $T(\\alpha_n), T(\\alpha_n)^\\perp$; this makes the regular-point tangent constraint a simple zero condition on the perpendicular components. Because the energy decouples into separate integrals when $t_n$ and $\\alpha_n$ are fixed, the optimization is an alternating scheme: local searches over $t_n$ and $\\alpha_n$, with per-interval gradient descent for the four Bézier parameters initialized by the linear least-squares estimate of equation (8), and the distance to the boundary evaluated through an efficient distance computation.","core_discovery":"The central claim is that minimizing the energy functional $$E(\\{t_n,\\alpha_n,\\lambda_n,\\gamma_n,\\beta_n,\\delta_n\\}) = \\sum_{n=1}^N \\$int_0^{{L_n}}$ (d_C(B_n(s)) + w_n)\\,\\|B_n'(s)\\|\\,ds$$ over a collection of cubic Bézier curves drives those curves onto the silhouette boundary $C(t)$. Here $d_C(\\bar x)$ is the Euclidean distance from a point $\\bar x$ to the boundary, and the parameters include the endpoint positions $t_n$, their tangent orientation angles $\\alpha_n$ at regular points, and the Bézier control parameters $\\lambda_n,\\gamma_n,\\beta_n,\\delta_n$. By writing the control points in terms of the tangent vectors $T(\\alpha_n)$ and its perpendicular, regular points are forced to meet the prescribed tangent by simply setting $\\gamma_n=\\delta_{n-1}=0$. The paper reports that alternating minimization of this energy, updating each $t_n$ within a small window and each $\\alpha_n$ within $\\pm 4$ degrees while solving per-interval gradient descents, reduces the average boundary-to-curve distance by 15.46 to 54.10 percent across the three silhouettes and three initializers.","pith_inferences":["Because the energy is non-convex, the reported gains likely depend on the starting vectorization; in broader use, feeding the model several initial guesses and keeping the lowest-energy output would probably improve robustness.","The energy is asymmetric, measuring only the distance from the Bézier curves to the boundary; the reported $d(C,B)$ improvements suggest the asymmetry is mild on these shapes, but a symmetric variant would be safer for shapes with long thin protrusions.","The method should extend to open curves and to multiple connected components by treating each Jordan curve independently, as the paper notes; a natural test is whether the gains persist on silhouettes with fine detail or noise."],"forward_implications":["Any vectorization method can be used as the initializer; in the reported experiments, Inkscape, Adobe Illustrator, and the curvature baseline all improved after the active contour refinement.","Both one-sided distances $d(B,C)$ and $d(C,B)$ decrease in the reported experiments, so the refined curves both lie close to the boundary and adequately cover it.","Raising $w_n$ on a particular Bézier segment shortens that segment, giving a built-in regularity control for removing undesirable irregularities.","Per-interval decoupling of the parameter estimation keeps the optimization cost roughly linear in the number of Bézier segments."],"supporting_citations":[{"why":"Supplies the geodesic active contour energy $E(B)=\\int g(B(s))\\|B'(s)\\|\\,ds$ that the proposed model minimizes.","marker":"[63]"},{"why":"Supplies the choice $g(\\bar x)=d_C(\\bar x)+w$ used as the weight in the active contour energy.","marker":"[66]"},{"why":"Provides the efficient algorithm used to compute the distance function $d_C$ from any point to the silhouette boundary.","marker":"[69]"},{"why":"Describes the Potrace algorithm underlying the Inkscape vectorization that serves as one of the initial guesses and comparison baselines.","marker":"[28]"},{"why":"Derives the linear least-squares system (8) used to initialize the Bézier curve parameters before gradient descent.","marker":"[31]"}],"fun_headline_variants":["Active contour cuts silhouette error up to 54%","Bézier active contour trims vectorization error 15-54%","Active contour refines Bézier traces from major tools","Silhouette vectorization improved via active contour model","Tighter Bézier curves with active contour energy minimization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the alternating minimization of the non-convex energy (6) reliably reaches a useful local minimum from whatever initial vectorization is supplied, so that the reported 15 to 54 percent improvement carries over to silhouettes and initial guesses beyond the three tested.","fun_headline_variants_meta":{"raw":{"variants":["Active contour cuts silhouette error up to 54%","Bézier active contour trims vectorization error 15-54%","Active contour refines Bézier traces from major tools","Silhouette vectorization improved via active contour model","Tighter Bézier curves with active contour energy minimization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000157,"raw_usage":{"total_tokens":1238,"prompt_tokens":977,"completion_tokens":261,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":180}},"tokens_in":593,"tokens_out":261,"duration_ms":3169,"temperature":1.0,"reasoning_tokens":180,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:11:47.065990+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the active contour model on a diverse set of silhouettes using deliberately poor initial guesses and measure both $d(B,C)$ and $d(C,B)$; if for some reasonable initial guess the minimization stalls in a local minimum that leaves large uncovered boundary stretches, or if the average distance increases relative to the initial vectorization, the claim that the method significantly improves any vectorization would fail.","supporting_citations":[{"cited_title":"International Journal of Computer Vision22, 61–79 (1997)","cited_arxiv_id":null,"evidence_quote":"Supplies the geodesic active contour energy $E(B)=\\int g(B(s))\\|B'(s)\\|\\,ds$ that the proposed model minimizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the choice $g(\\bar x)=d_C(\\bar x)+w$ used as the weight in the active contour energy."},{"cited_title":"IPOL Journal : Image Processing On Line11, 18–36 (2021) Appendix","cited_arxiv_id":null,"evidence_quote":"Provides the efficient algorithm used to compute the distance function $d_C$ from any point to the silhouette boundary."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the Potrace algorithm underlying the Inkscape vectorization that serves as one of the initial guesses and comparison baselines."},{"cited_title":"Journal of Mathematical Imaging and Vision 64, 41–56 (2022)","cited_arxiv_id":null,"evidence_quote":"Derives the linear least-squares system (8) used to initialize the Bézier curve parameters before gradient descent."}],"review_version":1}