{"id":"54f2c02e-098b-4bd4-9de1-a42cfe1be2d5","arxiv_id":"2411.08592","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A variational morphological skeleton prior, smoothed via log-sum-exp and unrolled into SAM's mask decoder, improves slender object segmentation in remote sensing images by roughly 1-2 F1 points.","lead":"The authors add a learnable morphological skeleton prior to the Segment Anything Model (SAM) by unrolling a variational optimization algorithm into a network module called MorSP. On building, road, and water segmentation benchmarks, SAM-MorSP improves F1 scores by about 1 to 2 points over fine-tuned SAM, with larger gains under added image noise.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The MorSP forward pass is a 20-iteration heuristic with one-step q/w updates and no convergence analysis; without a check that it approaches the minimizer of Eq. (10), the reported gains are not shown to come from the skeleton prior rather than added trainable capacity.","rationale":"The paper's headline contribution is not just a new module but a claim that the module is a faithful, interpretable unrolling of the variational skeleton-prior model in Eq. (10). The forward pass stops after T=20 iterations and uses single projected-gradient steps for q and w (Section III-B, Algorithm 1); no theorem, monotonicity argument, or numerical convergence study connects this finite network to the solution of the optimization problem. I selected this as the load-bearing concern because the reported accuracy gains are small (roughly 0.8-1.5 F1 points on clean sets) and are attributed specifically to the variational prior. Without a demonstration that the unrolled map approaches the model's minimizer, the most parsimonious explanation of the gains is the extra trainable freedom in the mask decoder token, learnable hyperparameters, and the auxiliary cl-dice loss. The check I propose directly measures whether T=20 iterations of the declared algorithm converge toward the true minimizer on representative inputs; if they do, the concern is retired. This is consistent with the reader's weakest assumption, so the CONDITIONAL verdict should stand.","tokens_in":16797,"tokens_out":7822,"duration_ms":76941,"concrete_test":"Run Algorithm 1 standalone (not inside SAM) on a fixed set of segmentation features o taken from SAM-fine and known skeleton targets S^alpha(g), with T=20 and the Table I hyperparameters. Compute a reference minimizer of Eq. (10) with a convergent first-order or ADMM solver run to tight tolerance. Compare u_20 with the reference in terms of relative objective gap and skeleton discrepancy ||S^alpha(u_20) - S^alpha(g)||_2. If u_20 is far from the reference, or if increasing T does not monotonically reduce the objective and skeleton discrepancy, the unrolled module is not implementing the declared variational model, and the reported gains cannot be attributed to the skeleton prior rather than to added trainable capacity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SAM-MorSP's advantage comes from a learnable morphological skeleton prior enforced by solving the variational model Eq. (10). But the MorSP forward pass (Algorithm 1, Section III-C) is not shown to solve that model: the q and w updates are single approximate gradient steps, and the loop is cut at T=20 'to save computational sources'. No convergence analysis, monotonicity result, or numerical comparison with a converged solution is provided. Thus the output u_T may be far from the minimizer, and the module may behave as a generic learned refinement layer. Since the measured gains over SAM-fine are small (F1 +0.84 to +1.45 on clean sets, with precision drops on LAIS building and water) and no error bars are given, the empirical tables do not isolate the skeleton prior as the source of improvement; added capacity in the token and learnable hyperparameters, plus the cl-dice loss, are equally consistent with the numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SAM-MorSP, an extension of the Segment Anything Model that integrates a learnable morphological skeleton prior into the segmentation decoder. The authors derive a smooth approximation of morphological dilation/erosion and of the morphological skeleton, use it to build the variational segmentation model in Eq. (10) with an entropy term, a threshold-dynamics regularization, and a skeleton cost, and then unroll an alternating optimization scheme into a trainable MorSP module with a soft skeleton token. The module replaces SAM's final Heaviside decoding step. Experiments on building, road, and water datasets compare SAM-MorSP with SAM-fine and SAM-cl, reporting improved F1 and recall in most settings, including under Gaussian noise, and an additional experiment integrates MorSP into the BEDSN backbone.","tokens_in":17104,"tokens_out":6698,"duration_ms":64140,"significance":"If the claims hold, the paper offers a principled way to inject a topological/skeleton prior into a foundation model through a variational energy rather than only through a loss function. The derivations in Propositions 1-2 provide a clean log-sum-exp representation of morphological operators, and the proposed module is architecture-agnostic, as demonstrated by the BEDSN experiment. The noise-robustness study is a useful addition, and the comparison with a cl-dice-loss baseline helps separate the effect of the variational module from the skeleton loss. However, the interpretability claim depends on two points that are not established: the smoothness of the derived skeleton variation, and the claim that the truncated, approximate unrolled iterations actually solve the variational model. The empirical evidence also lacks error bars and has precision drops on several datasets, so the significance is conditional on these points being resolved.","major_comments":[{"comment":"Proposition 3 expresses the variation of S^α(u) using the Heaviside step function H, and S^α itself is defined through ReLU projections. Consequently, S^α is not smooth in u, and the displayed formula is not the derivative of a smooth function; this contradicts the paper's assertion that the proposed skeletal representation is smooth and therefore suitable for gradient backpropagation. Since δC/δu in Eq. (9) and the w-update in Algorithm 1 depend on this formula, the authors should either reinterpret the expression as a subgradient/formal derivative, or replace the ReLU and Heaviside functions with genuinely smooth approximations and re-derive the variation.","section":"Section III-A, Proposition 3"},{"comment":"The paper states that the q and w subproblems are solved 'approximately' with one iteration each and that T = 20 layers are taken 'to save computational sources,' but it provides no convergence, monotonicity, or fixed-point analysis for the alternating scheme, and no numerical comparison with a converged solution of Eq. (10). As written, the MorSP forward pass may be far from the minimizer of the variational model, so the reported gains cannot be confidently attributed to the skeleton prior rather than to added trainable capacity in the soft skeleton token, the learnable hyperparameters, or the cl-dice loss. I ask for either a convergence/consistency analysis of Algorithm 1 or a numerical study reporting objective values and segmentation outputs for larger T and for a converged solution.","section":"Section III-B and Algorithm 1"},{"comment":"The headline claim of 'superior performance across all datasets, with an approximate 2% improvement in recall compared to SAM-fine' is not uniformly supported by the tables: on clean test sets, F1 gains over SAM-fine range from +0.79 to +1.45, recall changes range from +0.17 to +5.75, and precision drops on LAIS building (83.39 to 80.10), Massachusetts road (79.27 to 77.60), and WHU-OPT-SAR water (79.22 to 72.03). No error bars or statistical significance tests are reported, and Table V does not state whether the λl selection is made on the validation or test split; if the test set is used for hyperparameter selection, the reported numbers are optimistic. Please provide multi-run statistics and clarify or avoid test-set selection.","section":"Section V-C and Tables III-VII"}],"minor_comments":[{"comment":"The footnote says the three models have the same number of learnable parameters, but SAM-MorSP adds a soft skeleton token and learnable hyperparameters (γ, λ, α, η, ι) in addition to the modified decoder; please report the exact learnable-parameter count for each baseline.","section":"Table III footnote"},{"comment":"There are several typos and grammatical errors, e.g., 'divid' should be 'divide', 'searve' should be 'serve', 'konwledge' should be 'knowledge', 'segementing' should be 'segmenting', and 'an 32GB' should be 'a 32GB'; a careful proofreading pass is needed.","section":"Section II-A"},{"comment":"The loss-weight experiment in Table V is reported on the Massachusetts road dataset, but the text does not state whether these numbers are on the validation set or the test set; this should be clarified to assess whether the hyperparameter was selected legitimately.","section":"Section V-D"},{"comment":"The computational-cost table reports 'Gflops' and inference time, but the hardware and measurement details are minimal; please state the GPU model, batch size, and whether inference time is measured on a single 1024×1024 image.","section":"Section V-E"},{"comment":"The BEDSN experiment is described briefly without training details or error bars; adding the same experimental protocol as the SAM experiments would make the generalization claim more convincing.","section":"Section VI"}],"recommendation":"major_revision","confidential_remarks":"The paper presents an attractive way to combine variational morphology with a foundation model, and the experiments cover a reasonable set of remote sensing tasks. The main risk is that the mathematical-interpretability claim is currently stronger than what is demonstrated: the 'smooth' skeleton derivative contains a Heaviside term, and the truncated one-step alternating scheme is not connected to the minimizer of Eq. (10). If the authors can address these points by a formal subgradient statement or smooth approximation, and by a convergence/consistency study plus error bars, the paper would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the SAM-MorSP paper. The genuinely new thing here is a smooth, differentiable version of the morphological skeleton, built with log-sum-exp and a Fenchel dual representation, then folded into a variational segmentation model that gets unrolled into a network module (MorSP). Integrating that module into SAM via an extra skeleton token and a modified final sigmoid is clean and original. The control against SAM-cl with the same loss and same parameter count is a fair test, and the BEDSN experiment suggests the module transfers to other backbones.\n\nThe soft spots are real but not fatal. Proposition 3 calls the variation smooth while using a Heaviside step function, which undercuts the smoothness claim. More importantly, the unrolled forward pass is a heuristic: the q and w subproblems each take one approximate gradient step, the loop is cut at T=20 without a convergence check, and no numerical evidence shows that u_T is close to a minimizer of Eq. (10). So the skeleton prior, the module's stated mechanism, is not convincingly isolated from added trainable capacity. The empirical tables reinforce this worry: gains over SAM-fine are about 0.8–1.5 F1 points, with precision drops on LAIS building/water, and the loss weight λl was chosen on the test set. No error bars either.\n\nOn balance, the idea is sound and the paper is worth a serious referee. I would ask the authors to add convergence analysis or at least a numerical comparison with a fully converged solution, report error bars across runs, and move hyperparameter selection off the test set. With those changes, the central claim about the skeleton prior driving the improvement would be much stronger.\n\nMy take: send it to peer review. Not a desk reject.","headline":"A clever variational unrolling of a smooth morphological skeleton into SAM, with consistent but modest gains whose origin in the skeleton prior is not yet convincingly established.","tokens_in":17586,"tokens_out":1952,"would_cite":true,"duration_ms":19318,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68U10","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A learnable morphological skeleton prior embedded in SAM's decoder lifts recall by about 2 percent on slender remote-sensing objects.","keywords":["semantic segmentation","remote sensing","morphological skeleton","Segment Anything Model","variational method","soft threshold dynamics","algorithm unrolling","slender object segmentation"],"falsifier":"A reader could settle the mechanism by freezing all SAM-MorSP weights and replacing the learned skeleton token with a constant or with the skeleton of a different object class on the Massachusetts road test set: if road recall stays near the reported 74.33 while the skeleton input carries no road structure, the claimed skeleton-prior effect is not what produces the gain.","tokens_in":16618,"feed_emoji":"🛰️","tokens_out":5567,"duration_ms":50693,"temperature":0.7,"pith_summary":"The paper tries to show that the fine-detail loss seen in encoder-decoder segmentation models can be repaired by injecting a morphological skeleton prior at the decoding stage, rather than only into the loss function. It constructs a differentiable approximation of the morphological skeleton and a matching variational segmentation model, then unrolls the solver into a module called MorSP that plugs into SAM. The reported result is consistent gains on building, road, and water datasets, roughly 2 percent higher recall than SAM fine-tuned normally, plus better robustness to Gaussian noise. If true, this gives a mathematically interpretable way to keep structural priors active during prediction, not just during training.","feed_headline":"Skeleton prior in SAM lifts slender-object recall by ~2%","feed_subtitle":"Adding a learnable morphological skeleton module to SAM's decoder preserves roads, buildings, and water in remote sensing images.","key_machinery":"The load-bearing device is the smooth morphological skeleton $S^\\alpha(u)$, obtained by replacing the non-differentiable max and min of dilation and erosion with log-sum-exp soft operators, together with a variational segmentation energy that appends a skeleton-matching cost $C(u)=\\frac{1}{2}\\int_\\Omega |S^\\alpha(u)(x)-S^\\alpha(g)(x)|^2\\,dx$ to the soft threshold dynamics model. Operator splitting and $L^1$ duality turn the minimization into three subproblems, the $u$-update of which is a sigmoid; unrolling roughly twenty iterations of that update, with a learnable skeleton token standing in for the unknown ground-truth skeleton, gives the MorSP module.","core_discovery":"The paper's central claim is that a smooth, learnable version of the morphological skeleton can be written as a variational regularizer and unrolled into a neural module, and that inserting this module into SAM's mask decoder makes the network preserve slender and topologically complex structures in remote sensing images. The integration works by adding a skeleton token to the mask decoder and replacing SAM's final thresholding with a learnable sigmoid solve of a variational model. On the reported building, road, and water benchmarks, SAM-MorSP achieves the best F1, IoU, precision, and recall among the compared fine-tuning strategies, with about a 2 percent recall gain over SAM fine-tuned without the module, and it degrades less under added Gaussian noise.","pith_inferences":["A direct test the paper does not report is ablating the skeleton token: if feeding a constant or random token preserves the reported recall gains, then the skeleton prior itself is not carrying the improvement.","The same module could transfer to non-remote-sensing line-structure tasks such as retinal vessel or power-line segmentation, where connectivity matters more than boundary accuracy; the paper only evaluates buildings, roads, and water.","Because the skeleton is learned from a token inside the decoder, it could in principle be conditioned on the prompt (point or box) to emphasize different topological structures, an extension the paper does not explore.","The paper offers no convergence analysis for its one-step approximate subproblem updates, so a controlled experiment separating the skeleton branch from the extra learned parameters would clarify whether the gains come from the variational prior or simply from more capacity."],"forward_implications":["SAM-MorSP should segment narrow and elongated objects such as road intersections, roundabouts, and slender rivers more completely than SAM fine-tuned with plain cross-entropy or skeleton loss.","Because MorSP operates at the final decoding layer, it can be attached to other segmentation backbones; the paper demonstrates this with BEDSN and reports improved F1 and IoU on building and road datasets.","The model should degrade more gracefully under Gaussian noise than loss-only skeleton supervision, since the soft threshold dynamics regularization and smooth morphological operators act during inference as well as training.","Fine-tuning only about 4 percent of SAM ViT-B's parameters, through LoRA and MorSP, is enough to exceed full-decoder fine-tuning on the tested remote sensing benchmarks."],"supporting_citations":[{"why":"Supplies the Segment Anything Model backbone whose mask decoder is modified.","marker":"[15]"},{"why":"Provides the Soft Threshold Dynamics variational segmentation model that the proposed energy extends.","marker":"[27]"},{"why":"Defines the cl-dice skeleton loss used both for the SAM-cl baseline and as part of the training loss.","marker":"[10]"},{"why":"Supplies the BEDSN backbone used in the second experiment to show MorSP is backbone-agnostic.","marker":"[28]"},{"why":"Supplies the LoRA fine-tuning strategy used to adapt SAM's frozen image encoder cheaply.","marker":"[47]"},{"why":"Supplies the aerial building dataset used for the WHU building experiments.","marker":"[53]"},{"why":"Supplies the LAIS dataset with noisy building and road labels used for generalization testing.","marker":"[54]"},{"why":"Supplies the Massachusetts road dataset used for road segmentation and noise robustness tests.","marker":"[55]"},{"why":"Supplies the WHU-OPT-SAR optical images used for water segmentation.","marker":"[56]"}],"fun_headline_variants":["Morphological skeleton token boosts SAM on slender objects","SAM + learnable skeleton prior: +2% recall on thin structures","Variational skeleton module sharpens SAM for roads and buildings","Skeleton-aware SAM outdoes vanilla on slender remote sensing","Learnable skeleton prior lifts SAM's slender-object segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gain is credited to the skeleton regularizer, but the 20-step approximate solver is never shown to converge to the variational model's minimizer, so the improvement could in principle come from the extra learned parameters rather than from the skeleton prior.","fun_headline_variants_meta":{"raw":{"variants":["Morphological skeleton token boosts SAM on slender objects","SAM + learnable skeleton prior: +2% recall on thin structures","Variational skeleton module sharpens SAM for roads and buildings","Skeleton-aware SAM outdoes vanilla on slender remote sensing","Learnable skeleton prior lifts SAM's slender-object segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000532,"raw_usage":{"total_tokens":2536,"prompt_tokens":893,"completion_tokens":1643,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":1561}},"tokens_in":509,"tokens_out":1643,"duration_ms":10623,"temperature":1.0,"reasoning_tokens":1561,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:31:26.464747+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle the mechanism by freezing all SAM-MorSP weights and replacing the learned skeleton token with a constant or with the skeleton of a different object class on the Massachusetts road test set: if road recall stays near the reported 74.33 while the skeleton input carries no road structure, the claimed skeleton-prior effect is not what produces the gain.","supporting_citations":[{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"Supplies the Segment Anything Model backbone whose mask decoder is modified."},{"cited_title":"Deep convolutional neural networks with spatial regularization, volume and star-shape priors for image segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the Soft Threshold Dynamics variational segmentation model that the proposed energy extends."},{"cited_title":"clDice-a novel topology-preserving loss function for tubular structure segmentation,","cited_arxiv_id":null,"evidence_quote":"Defines the cl-dice skeleton loss used both for the SAM-cl baseline and as part of the training loss."},{"cited_title":"Boundary- enhanced dual-stream network for semantic segmentation of high- resolution remote sensing images,","cited_arxiv_id":null,"evidence_quote":"Supplies the BEDSN backbone used in the second experiment to show MorSP is backbone-agnostic."},{"cited_title":"Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set,","cited_arxiv_id":null,"evidence_quote":"Supplies the aerial building dataset used for the WHU building experiments."},{"cited_title":"Learning aerial image segmentation from online maps,","cited_arxiv_id":null,"evidence_quote":"Supplies the LAIS dataset with noisy building and road labels used for generalization testing."},{"cited_title":"Machine learning for aerial image labeling,","cited_arxiv_id":null,"evidence_quote":"Supplies the Massachusetts road dataset used for road segmentation and noise robustness tests."},{"cited_title":"MCANet: A joint semantic segmentation framework of optical and SAR images for land use classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the WHU-OPT-SAR optical images used for water segmentation."}],"review_version":1}