{"id":"158974c2-5d32-48a1-9cce-857d51dfb2b8","arxiv_id":"2607.19669","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A weakly supervised segmentation loss combining a simplex-constrained Potts model with an RKHS-based label extension improves scribble-trained networks over partial cross-entropy baselines.","lead":"This paper presents a segmentation loss for scribble-supervised deep networks, built from a Potts model with a smooth perimeter term and a fuzzy membership function obtained by kernel extension of sparse labels. On the CSSD test set it reports about 3 mIoU points over partial cross-entropy baselines and within 3.2 points of fully supervised training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Energy (6)/(9) is concave, not convex: the smooth perimeter term expands to a linear term minus ⟨v_k, Gσ*v_k⟩, so the claimed convexity and the Appendix C convergence justification fail.","rationale":"The reader's stated weakest assumption was the reliability of the RKHS fuzzy membership u, but the reader's rationale also identified the convexity error. I regard the convexity error as the most load-bearing concern because it undermines a formally stated central claim—'convex, smooth energy'—and the convergence justification in Appendix C. The empirical comparisons may still be valid, so a conditional accept remains appropriate: the theoretical framing must be corrected before the paper can be accepted as written. The Table 2 unnamed row is a second, also concrete, internal inconsistency that strengthens the need for revision. My recommendation is therefore unchanged from the reader's CONDITIONAL verdict.","tokens_in":23769,"tokens_out":6310,"duration_ms":75219,"concrete_test":"Compute the second variation of the discretized loss (9) with respect to a feasible network-output direction. For a single image and fixed u, take a perturbation δN with Σ_k δN_k = 0 and compute d²/dε² L(N + εδN) at ε=0. Because the fidelity term is linear, this reduces to −2λ Σ_k ⟨δN_k, Ĝ * δN_k⟩, which is non-positive for any δN. If this computation confirms negative semidefiniteness, the convexity claim is refuted. An even simpler two-pixel check: set u_1 = u_2 = 0.5, λ = 1, Ĝ = I, K = 2, and compare E at v = (1,0) and at the midpoint v = (0.5,0.5); the midpoint energy is larger, violating convexity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central theoretical claim—that the energy in (6)/(9) is convex—is false. Expanding the smooth perimeter term gives ∫(1−v_k)(Gσ*v_k)dx = ∫(Gσ*1)v_k dx − ∫ v_k(Gσ*v_k)dx. The first term is linear in v_k; the second is −⟨v_k, Gσ v_k⟩, which is concave because Gaussian convolution is positive semidefinite. In the discrete loss (9), the same expansion holds with Ĝ positive definite, so the Hessian with respect to Nθ,k is −2λĜ (up to simplex coupling), negative semidefinite. Thus the energy is concave, not convex, in both the segmentation variable v and the network output. This directly contradicts the abstract's 'convex, smooth energy functional' and the conclusion's restatement. It also invalidates the explicit justification in Appendix C, where the affine-majorization update and the cited convergence results [30, Prop. 2] and [32, Thm. 2.2] are invoked because the energy is convex. The empirical gains may still hold, but the theoretical guarantee that is a principal contribution of the paper is not supported. A secondary but concrete inconsistency: Table 2 contains an unnamed row with mIoU 75.76, higher than RKHS's 75.39, contradicting the text that 'none of them outperforms our RKHS'; this row appears to be PCE+CV at λ=1, and the claim needs correction regardless of formatting.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified variational framework for weakly supervised image segmentation. Sparse scribble labels are extended to a dense fuzzy membership function via an RKHS function-extension problem; this membership is then fed into a simplex-constrained Potts model with a smooth nonlocal perimeter surrogate. The resulting energy is proposed both as an iterative single-image segmentation model, solved by threshold dynamics, and as a training loss for a UNet. Experiments on ECSSD and PASCAL VOC report improvements over partial cross-entropy and several regularized baselines, together with analyses of the training effects. The central theoretical claim, repeated in the abstract and conclusion, is that the energy is convex and smooth and that this convexity justifies the optimization algorithms.","tokens_in":24185,"tokens_out":13128,"duration_ms":169908,"significance":"The paper's overall direction is valuable: converting sparse labels into dense soft supervision through an RKHS extension is a principled and computationally efficient idea, and the reported experiments are extensive, including ablations, failure cases, and a candid discussion of limitations. The single-image experiments cover noise, illumination bias, and water splashes, and the paper is transparent about parameter sensitivity. If the empirical claims are reproducible, the approach is a useful addition to weakly supervised segmentation. However, the advertised convexity of the energy is false, and the convergence justification for the threshold-dynamics solver is therefore unsupported as written. The empirical comparison also contains an unlabeled table row that contradicts a stated outperformance claim. These issues are central to the paper's contribution and require substantive correction.","major_comments":[{"comment":"The energy is not convex. Expanding the smooth perimeter term in (6) gives ∫(1−v_k)(Gσ∗v_k)dx = ∫v_k dx − ∫v_k(Gσ∗v_k)dx. Since Gaussian convolution is positive semidefinite, the second term is a negative quadratic form, so E is concave in v; the same holds for the loss in (9), where the Hessian with respect to Nθ,k contains −2λĜ. This contradicts the abstract's 'convex, smooth energy functional' and the conclusion. It also invalidates the Appendix C justification of Algorithm 3: the premise 'Since E is convex' is false, and the cited convergence results [30, Prop. 2] and [32, Thm. 2.2] cannot be invoked as stated. In addition, the printed update uses g_k^(t) = 1−2u_k + λGσ∗(1−v_k^(t)), but the first variation of the perimeter term is λ(1−2Gσ∗v_k), not λGσ∗(1−v_k); the missing factor of 2 changes the argmin. The authors should correct the characterization of the energy, the algorithm, an","section":"§2.3 (Eq. (6)), §2.4 (Eq. (9)), Appendix C"},{"comment":"Table 2 contains an unnamed row immediately after PCE+CV with λ=1, reporting mIoU 75.76 and mDice 85.76. These values are higher than the RKHS row (mIoU 75.39, mDice 85.52), directly contradicting the sentence 'none of them outperforms our RKHS.' The row should be labeled (presumably PCE+CV with λ=1) and the claim revised, or the row must be removed. This is not merely a typesetting issue: the comparison is central to the paper's claim of 'consistent improvements.'","section":"Table 2, §5.3.1"},{"comment":"The equivalence between the binary problem (4) and its relaxation (5) is asserted by citation to [36] and [32] rather than proved. Given the concavity issue above, the usual convex-relaxation reasoning does not apply, so the authors should state the precise conditions and provide a self-contained argument, or at least an exact statement of the cited lemma, so the reader can verify the equivalence.","section":"§2.2 (Eqs. (4)–(5))"}],"minor_comments":[{"comment":"The notation x, y is overloaded: they are both points in the image domain and (2R_k+1)×(2R_k+1) patches centered at those points. Please define the patch vectors explicitly, e.g., P(x) and P(y).","section":"Eq. (22)"},{"comment":"The text states that all models are trained for 8000 epochs. For 1400 training images and a 31M-parameter UNet, this is an unusually large number of full passes; please clarify whether 'epochs' means iterations/steps or true epochs, and report the actual training time.","section":"§5.1"},{"comment":"The last line of the proof contains a typographical artifact: 'ΨT he(1)γ ≥ · · · ≥ Ψγ(K)'. This should read 'Ψ(1)^γ ≥ ... ≥ Ψ(K)^γ' and should be corrected.","section":"Appendix B.2 (proof of Lemma 2)"},{"comment":"The caption says 'Smoothing effect of the perimeter term ... under different values of σ_s,0', but σ_s,0 is a kernel parameter of the RKHS membership function, not a parameter of the perimeter term. Please reword to avoid confusion.","section":"Figure 2 caption"},{"comment":"The NCut and KCut formulas use a bilateral kernel with a normalization constant C, but C is never specified. Please state how C is computed for the discrete implementation.","section":"Appendix E.3.3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has a strong empirical component and a useful algorithmic idea, but the advertised convexity is false and the iterative algorithm as printed does not match the first variation of the stated energy. The table inconsistency is also damaging to the empirical claims. I would be willing to review a revised version in which the theory is corrected or appropriately weakened, the algorithm is fixed, and the comparison claims are made consistent with the reported numbers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the empirical recipe is reasonable and the gains look real, but the paper's central theoretical claim is wrong. The loss combines a known smooth Potts/TD perimeter surrogate with an RKHS-based label extension and replaces a Moore-Penrose projection with a simplex threshold. That is a modest but concrete new combination, and the ECSSD/CSSD numbers are plausible: roughly +3 mIoU over PCE and within about 3 points of fully supervised CE. The qualitative training effects (denoising, edge refinement, boundary artifact removal) are well documented and worth replicating.\n\nThe problem is the convexity claim. Expanding the perimeter term in (6)/(9) gives a linear term minus <v_k, Gσ*v_k>. Gaussian convolution is positive semidefinite, so that quadratic term is concave. The Hessian with respect to the network output is -2λG-hat, negative semidefinite. So the energy is concave, not convex, and the appendix's 'Since E is convex' justification for the threshold dynamics convergence is not valid. This is not a typo or a minor misspeaking; the paper repeatedly advertises convexity as a principal contribution and uses it to justify the algorithm.\n\nThere are smaller issues too. Table 2 contains an unnamed row with mIoU 75.76, above RKHS's 75.39, contradicting the text that none of the baselines outperforms RKHS. That row appears to be PCE+CV at λ=1; the claim needs correction regardless of formatting. There are no error bars or multiple seeds. The RKHS membership u is precomputed with hand-chosen σI, σs, R, γ, and the whole method depends on u being a trustworthy proxy for true class membership; the cactus example in Fig. 4 shows the risk, but sensitivity is not quantified. The relaxed-to-original Potts equivalence is cited to the authors' own survey rather than proved.\n\nThe empirical method may still work as a heuristic, and the loss could be usable in practice. But the paper as written cannot be accepted with the convexity/convergence claims intact. It deserves a serious referee because the empirical study is substantial and the fix is localized: remove or substantially reframe the convexity claim, correct Table 2, add variance estimates, and discuss why threshold dynamics still behaves reasonably on a concave energy. Right now I would not cite it as a reliable theoretical result, but I would send it to review with the expectation of major revision.","headline":"Useful empirical package, broken convexity claim: the energy is concave, not convex, and the convergence justification in Appendix C does not hold.","tokens_in":24647,"tokens_out":2890,"would_cite":false,"duration_ms":34810,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Scribble labels become a fuzzy membership map via kernel interpolation, feeding one convex Potts energy used as both iterative solver and network loss, nearing full-supervision accuracy without ground-truth masks.","keywords":["weakly supervised segmentation","scribble labels","Potts model","RKHS function extension","fuzzy membership","threshold dynamics","convex variational loss","partial cross-entropy"],"falsifier":"Retrain the same network on the same scribbles but with deliberately corrupted u — for example, inflating the spatial kernel scale σ_s until labels bleed across true boundaries. The paper's account predicts the loss, and therefore the predictions, should follow u and degrade; if the network holds its mIoU anyway, the energy is not the mechanism carrying performance. A sharper quantitative version: on the 200-image CSSD test set, the RKHS-over-PCE gap should widen as the scribbles are made sparser; if the gap stays flat across scribble densities, the fuzzy membership is not doing the work the p","tokens_in":23681,"feed_emoji":"🎯","tokens_out":19452,"duration_ms":172819,"temperature":0.7,"pith_summary":"The paper claims that sparse scribble labels are enough to drive accurate image segmentation, provided they are first propagated across the whole image as a fuzzy membership function and then fed into a single convex variational energy. That energy — a simplex-constrained Potts model whose perimeter is a Gaussian-smoothed surrogate — can either be minimized directly by iterative thresholding for one image, or discretized into the training loss of a standard segmentation network. Trained this way, the network denoises the fuzzy map, sharpens object edges, and removes boundary artifacts; on the test set it beats partial cross-entropy (the usual scribble-supervision loss, which supervises only the labeled pixels) by roughly three mIoU points (mean intersection-over-union, a standard overlap metric), while staying within about three points of fully supervised training, all without seeing a ground-truth mask. If the claim holds, pixel-level annotations are not necessary to train competitive segmentation models — scribbles plus a well-chosen kernel suffice.","feed_headline":"Scribble-only training trails full supervision by just ~3 points","feed_subtitle":"Convex Potts-model energy with kernel-propagated scribbles beats partial cross-entropy, needs no ground-truth masks.","key_machinery":"Three components carry the argument. (1) The label-based fuzzy membership function u: sparse one-hot scribbles ψ on the labeled pixels D are extended to the whole domain by solving the RKHS linear system (K + mγI)a = ψ, giving Ψ(x) = Σ_i K(x,x_i) a_i; a Euclidean projection onto the probability simplex — implemented as a sorting-and-thresholding loop — turns Ψ into u, and the data fidelity takes the form f_k = 1 − 2u_k. The kernel is a product of a Gaussian in patch-intensity difference and a Gaussian in pixel distance, so its parameters control how far and how selectively labels propagate. (2) The smooth perimeter regularizer ('threshold dynamics', TD): λ Σ_k ∫ (1−v_k)(G_σ * v_k) dx, a Gaus","core_discovery":"The paper's central claim is that a single convex, smooth energy — a simplex-constrained Potts model over membership functions v, with data term driven by a fuzzy membership u and perimeter approximated by Gaussian convolution — serves both classical iterative segmentation and deep-network training. Sparse one-hot scribbles ψ are extended to the whole image by solving a regularized least-squares problem in a Reproducing Kernel Hilbert Space (RKHS, a function space whose kernel makes point evaluation continuous and enables smooth interpolation); the extension is projected onto the probability simplex by a thresholding loop to give u, and the data fidelity takes the form f_k = 1 − 2u_k. The pe","pith_inferences":["If the energy truly governs what the network learns, the three reported training effects should be visible as decreases in the corresponding terms of loss (9) over training; a direct check is to log the fidelity term and the TD term separately and verify that the TD term drops most when predictions over-segment.","The whole chain hinges on label propagation, and the paper concedes that kernel and weight parameters are picked empirically; this points to a testable upgrade — replacing the hand-tuned RKHS extension with a learnable propagation front end — which should raise the ceiling in exactly the cases the paper flags as weak (cactus glochids, an under-segmented pillar).","The paper notes that on the test set the TD term can erase true positives when predictions under-segment (mAcc drops by 1.23 points), which suggests a practical variant: make λ depend on the confidence or current perimeter of the prediction instead of fixing it globally.","Two load-bearing claims — convexity of the energy and equivalence between the relaxed and hard-constrained Potts problems — are inherited from previously published results cited by the paper rather than proved here; checking those proofs against the present formulation would settle whether the 'convex and smooth' guarantee is intact."],"forward_implications":["Scribble-only training becomes competitive with full supervision: the reported test gap to the fully supervised baseline is roughly 3 mIoU points, about half the gap that vanilla partial cross-entropy leaves.","One energy spans both regimes: the same convex problem solved by threshold dynamics for a single image becomes, after discretization, the network's training loss, so algorithmic insights transfer directly between classical and deep segmentation.","Because the energy is convex and smooth, training needs no alternating projections or hard label constraints, and the perimeter term is just a Gaussian convolution, keeping per-iteration cost low.","The TD term acts as a training regularizer, not merely a post-processing step: it penalizes over-segmentation and stabilizes the network against inventing spurious foreground regions that are absent from u.","The framework is generic in its regularizer — replacing the smooth perimeter with normalized-cut or compactness terms changes only one component, so the same RKHS label extension and loss derivation chain transfers to other spatial priors."],"fun_headline_variants":["Convex model turns scribbles into top segmentation","Unified loss: scribbles train net without masks","Kernel scribbles match full supervision within 3 pts","One convex energy: scribbles to segmentation","Scribble segmentation: no masks, near-full accuracy"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the fuzzy membership u, interpolated from scribbles by an RKHS whose kernel parameters are hand-chosen (the paper concedes they are empirical), faithfully represents the true class assignment on unlabeled pixels; wherever u is wrong, the data term 1−2u drives the solver and the network toward that wrong answer — a failure the paper itself exhibits with cactus glochids (Fig. 4) and an under-segmented pillar (Fig. 26).","fun_headline_variants_meta":{"raw":{"variants":["Convex model turns scribbles into top segmentation","Unified loss: scribbles train net without masks","Kernel scribbles match full supervision within 3 pts","One convex energy: scribbles to segmentation","Scribble segmentation: no masks, near-full accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000144,"raw_usage":{"total_tokens":961,"prompt_tokens":645,"completion_tokens":316,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":389,"completion_tokens_details":{"reasoning_tokens":255}},"tokens_in":389,"tokens_out":316,"duration_ms":3799,"temperature":1.0,"reasoning_tokens":255,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T12:02:11.239549+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same network on the same scribbles but with deliberately corrupted u — for example, inflating the spatial kernel scale σ_s until labels bleed across true boundaries. The paper's account predicts the loss, and therefore the predictions, should follow u and degrade; if the network holds its mIoU anyway, the energy is not the mechanism carrying performance. A sharper quantitative version: on the 200-image CSSD test set, the RKHS-over-PCE gap should widen as the scribbles are made sparser; if the gap stays flat across scribble densities, the fuzzy membership is not doing the work the p","supporting_citations":[],"review_version":1}