{"id":"86595ce0-5ba7-4be0-9622-0c183e46e8d0","arxiv_id":"2501.04846","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A Mamba-based edge detector achieves SOTA results on BSDS500 and produces multi-granularity edges on single-label datasets using an ELBO-supervised Gaussian decoder.","lead":"EDMB is an edge detection model that replaces transformers with Mamba state-space layers, reaching state-of-the-art ODS 0.837 on BSDS500. It also generates multi-granularity edges without requiring multi-label training data, which previous methods could not do.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The multi-granularity claim rests on an unvalidated learned variance; no evidence shows σ² encodes meaningful uncertainty, and the best γ schedule is selected on the test set.","rationale":"I read the paper in good faith: it proposes a concrete Mamba-based architecture, reports a competitive single-granularity ODS of 0.837 on BSDS500 without multi-scale testing or extra PASCAL-VOC, releases code, and provides ablations for each architectural component. Those parts are creditworthy. The load-bearing question is whether the multi-granularity contribution is real. The abstract claims multi-granularity ODS 0.851 and, more importantly, the first multi-granularity edge detector that works on single-label data. That claim is the main novelty over prior work like MuGE and UAED, which require multi-label annotations. The mechanism is the Gaussian decoder with ELBO loss and Eq. 14. For the claim to hold, the learned variance must contain usable uncertainty information. The paper's validation is qualitative (Figure 3/4) plus ODS numbers at γ schedules chosen on the test set (Table 5). There is no direct measurement of granularity, no comparison to human-perceived granularity levels, and no test of whether the variance is non-degenerate. The internal text/table inconsistency on BIPED (0.924 vs 0.906) is a separate reporting flaw that reinforces caution but is not the core scientific gap. My one concrete check—a constant-variance ablation—would directly settle whether σ² is doing the work. The reader's weakest assumption pointed at the same mechanism; I agree partially, with a slightly sharper emphasis on the missing validation and the test-set-selected γ. I do not think the paper should be rejected; the single-granularity result and code are useful, but the multi-granularity claim needs a concrete demonstration that the variance is informative. Hence the verdict stays CONDITIONAL/UNCHANGED.","tokens_in":14068,"tokens_out":6434,"duration_ms":58566,"concrete_test":"Train an ablation on BSDS500 in which the variance head is disabled and σ² is replaced by a constant (e.g., σ² ≡ 1) while keeping the ELBO loss and the Table 5 γ protocol unchanged; compare the resulting ODS and the spread of p_γ across γ ∈ {n/2−5, n∈{0..10}} to EDMB's reported 0.851/0.854. If the constant-variance variant produces a similar ODS spread and the same qualitative γ sensitivity, the learned σ² is not the source of granularity and the multi-granularity mechanism is unsupported; if the variant collapses to a single ODS, the mechanism is validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Sections 3.4–3.6 construct per-pixel Gaussian distributions N(μ, σ²) and generate multi-granularity edges via p_γ = μ + γ σ² (Eq. 14). The paper's central novelty—multi-granularity edges on single-label data—depends entirely on the learned σ² capturing meaningful per-pixel edge uncertainty. The paper never validates this. (i) No statistics on the learned σ²: is it spatially structured, or nearly constant? (ii) No correlation of σ² with per-pixel annotator disagreement on BSDS500, which is the natural target for uncertainty. (iii) No ablation replacing σ² with a constant or with a hand-crafted proxy. (iv) The headline multi-granularity ODS 0.851 is obtained by selecting the best γ schedule (Random n/2−5) in Table 5 directly on the test set, so it is unclear whether any benefit survives an honest validation split. If σ² is degenerate or unrelated to edge ambiguity, Eq. 14 is just a global bias, and the claimed granularity control—and the first-to-do-it-on-single-label claim—does not follow from the evidence presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EDMB, an edge detector built on vision Mamba, combining a global Mamba encoder, a fine-grained Mamba encoder, and a lightweight CNN high-resolution encoder. A decoder predicts per-pixel Gaussian distributions N(mu, sigma^2), supervised by an ELBO loss composed of a weighted cross-entropy term and a KL divergence to N(0,1). Multi-granularity edges are generated as p_gamma = mu + gamma*sigma^2. Experiments on BSDS500, NYUDv2, and BIPED report single-granularity ODS 0.837 and multi-granularity ODS 0.851 on BSDS500, with further multi-granularity results on single-label datasets. The paper claims to be the first Mamba-based edge detector and the first to produce multi-granularity edges without multi-label training data.","tokens_in":14290,"tokens_out":5798,"duration_ms":50904,"significance":"If the reported results are sound, EDMB is significant in two respects: it demonstrates that a Mamba-based backbone can reach the top of the BSDS500 leaderboard without multi-scale testing or extra PASCAL-VOC data, and it extends multi-granularity edge generation to single-label datasets, removing the multi-label-data requirement of prior work such as UAED and MuGE. The paper is also commendable for providing source code, an honest ablation showing that the proposed decoder alone does not improve single-granularity accuracy, and a correct closed-form KL divergence in Eq. (8). The main risk is that the central multi-granularity claim rests on an unvalidated learned variance and on a gamma schedule chosen directly from test-set performance; these issues need to be resolved before the significance can be fully assessed.","major_comments":[{"comment":"The text states that on BIPED \"EDMB* achieves 0.924 ODS, which is the new SOTA method and 0.7% higher than the second best DiffusionEdge,\" but Table 3 reports EDMB* ODS of 0.906 and DiffusionEdge ODS of 0.899 for BIPED. The two numbers (0.924 and 0.906) are inconsistent, and the claimed 0.7% improvement matches the 0.906 vs. 0.899 difference, not the 0.924 value. Please correct this inconsistency and, if 0.924 was obtained under a different evaluation setting, describe that setting explicitly.","section":"Section 4.3, Table 3"},{"comment":"The headline multi-granularity ODS of 0.851 on BSDS500 is obtained by selecting the gamma schedule \"Random n/2 - 5\" from Table 5, which reports test-set ODS values between 0.845 and 0.851 for different schedules. No held-out validation set is used to select gamma, so the reported improvement from single-granularity 0.837 to multi-granularity 0.851 may reflect test-set overfitting rather than a genuine granularity-control benefit. Please report how gamma was selected (e.g., on a validation split) and, ideally, the variance of the results over multiple runs.","section":"Section 3.6, Eq. (14), Table 5"},{"comment":"The multi-granularity mechanism assumes that the learned per-pixel variance sigma^2 encodes meaningful edge uncertainty, so that p_gamma = mu + gamma*sigma^2 produces genuine granularity variation. The paper does not provide any direct evidence for this: no statistics on the learned sigma^2 (e.g., whether it is spatially structured or nearly constant), no correlation with per-pixel annotator disagreement on BSDS500, and no ablation that replaces sigma^2 with a constant or a hand-crafted uncertainty proxy. If sigma^2 is degenerate, Eq. (14) reduces to a global bias and the claimed ability to control granularity on single-label data does not follow. Please add such validation analyses.","section":"Sections 3.4-3.6, Eq. (14)"},{"comment":"The ELBO loss depends on \"a sampling p of the distributions,\" but the paper does not state how p is sampled during training, e.g., whether the reparameterization trick is used, how many samples are drawn per iteration, or whether the sampling is performed on the full-resolution output. This implementation detail is essential for reproducing the training procedure and for the claim that Eq. (10) is a valid ELBO. Please clarify the sampling procedure in the implementation details.","section":"Section 3.5, Eq. (10)"}],"minor_comments":[{"comment":"The integral derivation of the KL divergence is malformed in the typeset text (missing integral signs and unclear exponent placement). The final closed form is correct, but the derivation should be typeset properly for readability.","section":"Eq. (8)"},{"comment":"The term \"sampling\" is used both for random draws from the Gaussian during training and for the deterministic operation in Eq. (14), which may confuse readers. Consider using \"granularity modulation\" or \"deterministic shifting\" for the inference-time operation.","section":"Section 3.6"},{"comment":"The text says EDMB* \"can match\" RankED on NYUDv2, but Table 3 shows EDMB* ODS 0.783 above RankED's 0.780; this could be phrased more precisely as a slight improvement.","section":"Section 4.3, Table 3"},{"comment":"The statement that ODS and OIS are improved by \"0.3%\" over DiffusionEdge should say \"0.3 percentage points\" (i.e., 0.003 in the reported metrics), to avoid ambiguity between relative and absolute improvement.","section":"Section 4.3"},{"comment":"The caption of Figure 3 appears to be duplicated, and some subfigure labels (e.g., (g), (m)) are repeated. Please clean up the figure caption and subfigure numbering.","section":"Figure 3"},{"comment":"The claim that EDMB is \"the first Mamba-based edge detector\" should be verified against concurrent work, since the field is rapidly evolving; if any prior Mamba-based edge detector exists, the claim should be adjusted.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The central issue is that the multi-granularity contribution is validated only by test-set-selected numbers and qualitative figures; the learned variance itself is not analyzed. The BIPED text/table inconsistency is a clear factual error that must be fixed. If the authors can provide evidence that sigma^2 is meaningful and use a validation-based gamma selection, the paper could become a solid contribution to the edge-detection literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the EDMB paper in full. The bottom line: it is a reasonable engineering paper that delivers a competitive Mamba-based edge detector and a clever way to get multi-granularity outputs from single-label supervision. The single-granularity numbers are believable. The multi-granularity story is the weakest part, and the paper under-validates it.\n\nWhat's actually new: as far as the cited literature goes, this is the first Mamba-based edge detector, and the first to generate multi-granularity edges without multi-label training. The design is sensible: a global Mamba encoder for context, a windowed fine-grained Mamba encoder for local detail, a light CNN branch for high-res features, and a Gaussian decoder supervised with an ELBO loss. The KL derivation is correct. They release code, which is a plus. On BSDS500 they hit 0.837 ODS without multi-scale test or extra data, which is competitive with DiffusionEdge at about a third of the parameters and half the FLOPs. That part holds up.\n\nThe soft spots are real. First, there is a concrete numerical contradiction: the text says EDMB* achieves 0.924 ODS on BIPED, but Table 3 reports 0.906. That's a big enough gap to warrant a correction before anything is final. Second, the headline multi-granularity result (0.851 on BSDS500) is obtained by trying several gamma schedules and picking the best one on the test set, with no validation split. That inflates the comparison with UAED*/MuGE*. Third, and most important, the entire multi-granularity mechanism rests on the learned per-pixel variance sigma^2 encoding meaningful edge uncertainty. The paper never shows that it does. There are no statistics on sigma^2, no correlation with annotator disagreement, no ablation with a constant or hand-crafted variance. Without that, p_gamma = mu + gamma sigma^2 could simply be a global bias, and the 'multi-granularity' claim is not yet established. The authors' own limitation note admits higher cost than CNN and no downstream-task study, which tempers the practical impact.\n\nFor a reader: the single-granularity system is worth knowing about if you work on efficient edge detection or SSM-based vision. The multi-granularity contribution needs stronger evidence before it should be built upon.\n\nMy recommendation: send it to peer review, but with clear requests—fix the BIPED inconsistency, report variance statistics and an ablation that replaces sigma^2, and use a proper validation split for selecting gamma. The paper deserves a serious referee, but it needs a revision.","headline":"Solid Mamba-based edge detector with competitive single-granularity results, but the multi-granularity claim rests on unvalidated learned variance and test-set-selected gamma.","tokens_in":14880,"tokens_out":3691,"would_cite":true,"duration_ms":35672,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Mamba-based edge detector reaches top BSDS500 accuracy while generating multi-granularity edges without multi-label data.","keywords":["edge detection","Mamba","state space model","multi-granularity edges","evidence lower bound","learnable Gaussian distributions","single-label learning","BSDS500"],"falsifier":"Train EDMB on BSDS500 and inspect the learned $\\sigma^2$: if the average variance on edge pixels is not larger than on non-edge pixels, or if it does not correlate with the spread of human annotations across the 4–9 ground-truth maps, then the learned distribution is not encoding edge uncertainty and the multi-granularity claim fails. A simpler check is to fix $\\gamma$ at several values and compute the mean distance between the resulting edge maps; near-zero distance means $\\gamma$ does not control granularity.","tokens_in":13830,"feed_emoji":"🖼️","tokens_out":7773,"duration_ms":65116,"temperature":0.7,"pith_summary":"This paper tries to show that a vision-Mamba backbone can replace the Transformer in edge detection without sacrificing accuracy, and that multi-granularity edges do not require multiple human label sets. It reports an edge detector, EDMB, that reaches an ODS F-measure of 0.837 on BSDS500 at single granularity and 0.851 when granularity is varied, using no multi-scale testing and no extra PASCAL-VOC data. The same detector works on single-label datasets NYUDv2 and BIPED, where prior multi-granularity methods could not be applied. If true, this makes high-quality edge detection cheaper and makes granularity control available for datasets annotated by only one person.","feed_headline":"Mamba edge detector tops BSDS500, adds granularity from one label","feed_subtitle":"It is the first Mamba-based edge detector at top accuracy, and it produces coarse-to-fine edges from single-label data.","key_machinery":"The load-bearing object is the learnable Gaussian distributions (LGD) decoder. It turns fused global, fine-grained, and high-resolution features into per-pixel parameters $(\\mu, \\sigma^2)$; the edge map is a sample from $\\mathcal{N}(\\mu, \\sigma^2)$, and granularity is controlled by the scalar formula $p_\\gamma = \\mu + \\gamma \\sigma^2$. The ELBO loss—a weighted cross-entropy term for the sample plus a KL term that regularizes each pixel distribution toward $\\mathcal{N}(0,1)$—is what lets one binary label supervise the variance, removing the need for multiple human annotations. Without this decoder, the network still detects edges but cannot produce multi-granularity outputs.","core_discovery":"EDMB's central claim is that Mamba's selective state-space model can serve as the core of a state-of-the-art edge detector when paired with a global-local architecture and a probabilistic decoder. The network uses a global Mamba encoder for long-range context, a fine-grained Mamba encoder over non-overlapping windows for local detail, and a lightweight CNN high-resolution encoder to preserve precise location information. A learnable Gaussian distributions decoder fuses these features to predict per-pixel means and variances, and multi-granularity edges are obtained by sampling $p_\\gamma = \\mu + \\gamma \\sigma^2$. The Evidence Lower Bound loss—weighted cross-entropy on a sampled edge map plus a KL divergence pushing each pixel distribution toward $\\mathcal{N}(0,1)$—supervises the distribution so that a single binary label is enough. The paper reports ODS 0.837/0.851 on BSDS500 without multi-scale testing or extra PASCAL-VOC data, and ODS 0.783 on NYUDv2 and 0.906 on BIPED for the multi-granularity version, claiming the first Mamba-based edge detector and the first multi-granularity edge results on single-label datasets.","pith_inferences":["If the learned variance truly reflects per-pixel edge uncertainty, the same $\\sigma^2$ could guide active re-annotation by pointing at pixels where annotators most disagree; the paper does not explore this use.","The $\\gamma$ schedule is selected after evaluating several schedules and reporting the best ODS on the test set; a stricter protocol that fixes $\\gamma$ on validation data would be needed to know whether the 0.851 multi-granularity number is robust.","Because the ELBO head resembles a variational autoencoder, a natural extension is to condition $\\gamma$ on a task embedding so downstream systems could request coarse or fine edges explicitly instead of tuning a scalar.","A promising test of the method's generality is to apply it to domains with deliberately coarse annotation styles, such as medical or remote-sensing boundaries, where multi-label training data is rarely available."],"forward_implications":["Mamba can replace Transformers in edge detection: on BSDS500, EDMB reports a higher ODS than DiffusionEdge while using roughly one-third the parameters and about half the GFLOPs.","Multi-granularity edge maps become available on single-label datasets such as NYUDv2 and BIPED, removing a previous dependency on multiple human annotations.","Granularity is controlled at inference time by one scalar $\\gamma$, so a single trained model can serve tasks that want coarse boundaries or fine detail without retraining.","The reported gains on BSDS500 are achieved without multi-scale testing or extra PASCAL-VOC data, so they come from the architecture and loss rather than test-time augmentation."],"supporting_citations":[{"why":"Supplies the bidirectional Vision Mamba backbone that EDMB adapts for the global and fine-grained encoders.","marker":"[55]"},{"why":"Provides the global-local encoder design, spatial feature transform, and multi-stage training strategy that EDMB follows.","marker":"[28]"},{"why":"Defines the multi-granularity edge setting and the $p_\\gamma = \\mu + \\gamma \\sigma^2$ granularity control that EDMB adopts.","marker":"[53]"},{"why":"Introduces learnable Gaussian distributions over edges for uncertainty-aware multi-label training, which EDMB reworks for single-label data.","marker":"[54]"},{"why":"Supplies the ELBO loss formulation that combines a reconstruction term with KL divergence to supervise the predicted distributions.","marker":"[21]"},{"why":"Provides the BSDS500 dataset and the evaluation protocol used for the main accuracy comparisons.","marker":"[1]"},{"why":"Is the strongest Transformer-based baseline on BSDS500 that EDMB claims to surpass in accuracy and efficiency.","marker":"[49]"},{"why":"Supplies the weighted cross-entropy loss and the mixed-label handling used in training and in the multi-label comparison.","marker":"[24]"},{"why":"Introduces the selective state-space model that underpins Mamba and motivates EDMB's efficiency claims.","marker":"[13]"}],"fun_headline_variants":["Mamba edge detector reaches top BSDS500, adds multi-granularity","First Mamba edge detector with coarse-to-fine edges from single label","EDMB: Mamba-based detector hits ODS 0.851 on BSDS500","Mamba encoder yields state-of-the-art edge detection and granularity","Multi-granularity edges from one label via Mamba and Gaussian decoder"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The multi-granularity capability rests on the assumption that a per-pixel Gaussian variance, learned from a single binary edge label through the ELBO loss, actually captures meaningful edge uncertainty; if the variance is degenerate or unrelated to annotation disagreement, the granularity slider $p_\\gamma = \\mu + \\gamma \\sigma^2$ produces no real granularity variation.","fun_headline_variants_meta":{"raw":{"variants":["Mamba edge detector reaches top BSDS500, adds multi-granularity","First Mamba edge detector with coarse-to-fine edges from single label","EDMB: Mamba-based detector hits ODS 0.851 on BSDS500","Mamba encoder yields state-of-the-art edge detection and granularity","Multi-granularity edges from one label via Mamba and Gaussian decoder"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000204,"raw_usage":{"total_tokens":1437,"prompt_tokens":1042,"completion_tokens":395,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":658,"completion_tokens_details":{"reasoning_tokens":293}},"tokens_in":658,"tokens_out":395,"duration_ms":3826,"temperature":1.0,"reasoning_tokens":293,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:23:38.969517+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train EDMB on BSDS500 and inspect the learned $\\sigma^2$: if the average variance on edge pixels is not larger than on non-edge pixels, or if it does not correlate with the spread of human annotations across the 4–9 ground-truth maps, then the learned distribution is not encoding edge uncertainty and the multi-granularity claim fails. A simpler check is to fix $\\gamma$ at several values and compute the mean distance between the resulting edge maps; near-zero distance means $\\gamma$ does not control granularity.","supporting_citations":[{"cited_title":"Edter: Edge detection with transformer","cited_arxiv_id":null,"evidence_quote":"Provides the global-local encoder design, spatial feature transform, and multi-stage training strategy that EDMB follows."},{"cited_title":"Muge: Multiple granularity edge detection","cited_arxiv_id":null,"evidence_quote":"Defines the multi-granularity edge setting and the $p_\\gamma = \\mu + \\gamma \\sigma^2$ granularity control that EDMB adopts."},{"cited_title":"The treasure beneath multiple an- notations: An uncertainty-aware edge detector","cited_arxiv_id":null,"evidence_quote":"Introduces learnable Gaussian distributions over edges for uncertainty-aware multi-label training, which EDMB reworks for single-label data."},{"cited_title":"Contour detection and hierarchical image seg- mentation","cited_arxiv_id":null,"evidence_quote":"Provides the BSDS500 dataset and the evaluation protocol used for the main accuracy comparisons."},{"cited_title":"Richer convolutional features for edge detection","cited_arxiv_id":null,"evidence_quote":"Supplies the weighted cross-entropy loss and the mixed-label handling used in training and in the multi-label comparison."}],"review_version":1}