{"id":"8c46a66e-b008-4a7c-bd00-86aad811d249","arxiv_id":"2501.18855","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A new fusion-based crack segmentation model using a frozen EdgeSAM encoder with gated attention reports state-of-the-art accuracy and strong zero-shot generalization on pavement datasets.","lead":"FlexiCrackNet combines a frozen EdgeSAM image encoder with a U-Net style decoder to segment road cracks. The authors report state-of-the-art accuracy and strong zero-shot generalization across three pavement crack datasets, though no code is released.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that frozen EdgeSAM features drive the improvement is untested: the comparison lacks a capacity-matched from-scratch U-Net and any SAM fine-tuning baseline, so the reported gains could come from added parameters and compute rather than from general feature transfer.","rationale":"The reader's conditional verdict is reasonable, and my stress-test sharpens one aspect of it. Even taking Tables I-III at face value, the experiments do not establish that frozen EdgeSAM features are the active ingredient. The only trained-from-scratch comparison is a small U-Net with 18.64M parameters and 101.99 GFLOPs, while the proposed model uses 25.63M parameters and 196.18 GFLOPs. The Concat ablation shows that adding EdgeSAM features without IGAM hurts performance, so the attention mechanism, not the features per se, drives the gain. This leaves two plausible readings: either the frozen features plus IGAM genuinely transfer general knowledge, or a larger-capacity model trained from scratch with an effective attention module would do just as well. The paper does not currently distinguish between these. A capacity-matched U-Net baseline is the cheapest check that would settle this concern. The missing SAM-Adapter or fine-tuned EdgeSAM baseline is also relevant because the paper's motivation explicitly argues against the pre-training plus fine-tuning paradigm, yet no such method is compared. The arbitrary-input-size claim is separately unsupported since all experiments are at 512x512, consistent with the reader's note. No code or error bars are provided, which further weakens verification. These issues do not invalidate the paper, but they justify keeping the conditional verdict and requiring the missing baselines and reproducibility artifacts.","tokens_in":15443,"tokens_out":7990,"duration_ms":92174,"concrete_test":"Train a U-Net with channel width expanded to match FlexiCrackNet's parameter count (~25.6M) and FLOPs (~196 GFLOPs) using the exact training protocol of Section IV-A, and report DeepCrack test and zero-shot CFD metrics. If the scaled U-Net matches or exceeds FlexiCrackNet's 82.88 F1 on DeepCrack and 54.48 F1 on CFD, the gains are attributable to capacity and compute rather than to frozen EdgeSAM features; if it does not, the general-feature-transfer mechanism receives direct support. This check should be run alongside a SAM-Adapter or fine-tuned EdgeSAM baseline under the same protocol to test the claimed advantage over the pre-training plus fine-tuning paradigm.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that frozen EdgeSAM encoder features provide general visual knowledge that transfers better than either trained-from-scratch encoders or fine-tuned SAM variants. The evidence does not isolate this mechanism. In Table I, FlexiCrackNet uses 25.63M parameters and 196.18 GFLOPs, while the U-Net baseline uses 18.64M parameters and 101.99 GFLOPs; the gain on DeepCrack is +3.24 F1 (Table III), an amount that could simply reflect about 37% more parameters and roughly 92% more FLOPs. The ablation that replaces IGAM with plain concatenation ('Ours (Concat)', Table III) degrades below U-Net, so the frozen SAM features are not helpful on their own; only the learned gating mechanism restores performance. No U-Net scaled to the same budget is evaluated, and no SAM-Adapter or fine-tuned EdgeSAM variant is compared, despite SAM-Adapter being cited in Section II-B. Thus the paper's core premise that general SAM features are superior to fine-tuning or from-scratch training remains untested. The arbitrary-resolution claim is also unsupported: all experiments use 512x512 inputs, so there is no evidence that the pipeline delivers 'diverse input resolutions' as claimed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FlexiCrackNet, a crack segmentation pipeline built on a U-Net encoder-decoder augmented with a frozen EdgeSAM encoder used as a generic feature extractor. The two feature streams are fused at five encoder stages by an information-interaction gated attention mechanism (IGAM) that generates attention masks for both general and crack-specific features, with a residual connection preserving the crack-specific stream. Training uses a combination of BCE and Dice losses. The method is evaluated on the DeepCrack training/test split and in a zero-shot protocol on the CFD and Crack500 datasets, comparing against twelve baselines and reporting F1, IoU, Dice, parameters, inference time, and GFLOPs. The authors report state-of-the-art results on DeepCrack and large zero-shot gains on CFD and Crack500, and claim flexibility with respect to input resolution and computational efficiency.","tokens_in":15692,"tokens_out":2884,"duration_ms":30310,"significance":"If the claims are confirmed, FlexiCrackNet would be a practically useful way to inject foundation-model priors into crack segmentation while avoiding the overhead of full SAM fine-tuning. The evaluation protocol is sound in its use of genuinely held-out datasets for zero-shot assessment, and the reported improvements are consistent across three datasets and multiple baselines. The authors also make the efficiency comparison transparent by reporting parameters, inference time, and GFLOPs. However, the central mechanistic claim that frozen EdgeSAM features are the source of the improvement is not isolated by the experiments, and the arbitrary-resolution claim is not tested. The lack of code release and single-run results additionally limit verification. The work is a reasonable engineering contribution, but the evidence currently supports a more modest claim than the one made in the abstract.","major_comments":[{"comment":"The central claim that frozen EdgeSAM general features drive the improvement is not isolated. In Table III, replacing IGAM with plain concatenation ('Ours (Concat)') yields DeepCrack F1 78.78, below the U-Net baseline's 79.64, so the frozen EdgeSAM features are not helpful on their own; only the learned IGAM restores performance. Since Table I shows that Ours uses 25.63M parameters and 196.18 GFLOPs versus U-Net's 18.64M parameters and 101.99 GFLOPs, the +3.24 F1 gain over U-Net could plausibly come from added capacity and compute rather than from general feature transfer. A capacity-matched U-Net with comparable parameters and FLOPs should be evaluated to separate these effects.","section":"§IV-F, Table III"},{"comment":"The paper argues that the proposed paradigm is preferable to the 'pre-training + fine-tuning' approach for SAM, but no SAM fine-tuning baseline is compared. SAM-Adapter is cited in Section II-B, and Zhou et al. [23] is described as adapting SAM for road crack segmentation, yet neither a SAM-Adapter nor a fine-tuned EdgeSAM variant appears in Tables I or II. Without such a comparison, the claimed advantage over the fine-tuning paradigm is unsubstantiated.","section":"§II-B, §IV-D"},{"comment":"The claim of flexibility to 'diverse input resolutions' and 'arbitrary sizes' is not experimentally supported. Section IV-A states that all input images are resized to 512×512 pixels, and all experiments in Tables I and II use that resolution. No experiment varies the input resolution, so the decoupling of EdgeSAM from its fixed 1024×1024 input is demonstrated only architecturally, not empirically.","section":"§I, §III-B, §IV-A"},{"comment":"All reported results are from single runs with no error bars or statistical significance tests. Some of the headline margins are small, for example 1.47 F1 over CrackMamba on DeepCrack and 2.45 F1 over CrackMamba on the Crack500 test set, so it is unclear whether these differences are robust to training variance. Reporting mean and standard deviation over at least three runs, or providing significance tests, would strengthen the central claim.","section":"§IV-D, Tables I and II"},{"comment":"The statement that the method introduces only 'minimal additional computational cost' is misleading when compared to the U-Net baseline. Table I shows inference time increases from 4.59 ms for U-Net to 22.09 ms for FlexiCrackNet, and GFLOPs increase from 101.99 to 196.18. The efficiency advantage is real only relative to heavier transformer-based or SAM-based models such as CrackMamba; the comparison against the from-scratch U-Net should be stated more carefully.","section":"§IV-D, 'Efficiency comparison'"}],"minor_comments":[{"comment":"The word 'Transfered' in the title should be spelled 'Transferred'.","section":"Title"},{"comment":"The dataset referred to as 'Cityforest' is commonly known as the Crack Forest Dataset (CFD); the name should be corrected for consistency with the literature.","section":"§IV-B"},{"comment":"The figure caption lists EMCAD twice, as both (c) and (e); the second instance appears to be a typo.","section":"Fig. 7"},{"comment":"There are two different DeepCrack references, [38] and [46], with different authors and years; the in-text citations should disambiguate which work is the dataset and which is the segmentation method.","section":"References"},{"comment":"The experimental configuration states that all models use 'a fixed loss function', but the paper does not specify whether this is the same BCE+Dice loss for all baselines or each baseline's original loss; clarifying this would strengthen the fairness of the comparison.","section":"§IV-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for the journal and the zero-shot evaluation protocol is a genuine strength. The main risk is that the central attribution to frozen SAM features is not isolated by the current ablations; adding a capacity-matched baseline and a SAM fine-tuning baseline is necessary before the claims can be accepted. I would also encourage the editor to ask for code release, given that the reported margins are moderate and no error bars are provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about FlexiCrackNet. First, it is a clean, well-run study: the authors take a frozen EdgeSAM encoder, fuse its multi-scale features into a U-Net with a gated attention module (IGAM), and report consistent gains across three public crack datasets, including a genuinely held-out zero-shot evaluation on CFD and Crack500. The improvements over a dozen baselines are substantial and the visualization is convincing. Second, despite the abstract, the paper does not actually show that frozen SAM features are what drive the gains. The ablation table is the tell: simply concatenating the EdgeSAM features to the U-Net encoder (Ours/Concat) makes things worse than the plain U-Net baseline on both DeepCrack and CFD. Only after adding IGAM, which itself adds a nontrivial number of parameters and FLOPs, does performance jump. So the result could just as plausibly come from the extra capacity of the gating module rather than from any 'general knowledge' in the frozen encoder. The authors never test a capacity-matched U-Net, nor do they compare against a fine-tuned SAM variant like SAM-Adapter, despite citing it. That is a genuine hole in the central claim.\n\nThe paper does some things well. The zero-shot protocol is sound, the datasets are appropriate, and the efficiency numbers on parameters and inference time are genuinely better than the prior crack-specific SOTA, even though GFLOPs are about the same as CrackMamba. The qualitative analysis of failure cases (shadow, blur, texture) is also thoughtful. But the soft spots are real: no error bars, single-run results, no code release, and the 'arbitrary input resolution' claim is unsupported because every experiment uses 512×512 inputs. These are not fatal flaws for an applied paper, but they do mean the headline contribution should be read as 'a fusion module that works well with this particular frozen encoder' rather than 'frozen SAM features are superior to fine-tuning or training from scratch.'\n\nWho is the audience? Practitioners in structural inspection and road maintenance who want a stronger crack segmentation baseline will get value from this pipeline. People looking for insight into how or why foundation-model features transfer will not find it here. I would send this to peer review—the engineering is solid and the zero-shot results are worth reporting—but I would ask the authors to add a capacity-matched U-Net and at least one fine-tuned SAM baseline, report variance across runs, and either demonstrate the resolution flexibility or drop that claim.","headline":"Solid engineering result with a good zero-shot protocol, but the paper's story that frozen SAM features are the magic ingredient is not backed by its own ablation—review it, but push for the missing baselines.","tokens_in":16253,"tokens_out":1918,"would_cite":false,"duration_ms":22079,"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":"FlexiCrackNet claims that a frozen EdgeSAM encoder, fused into a U-Net through gated attention, beats twelve crack-segmentation methods on DeepCrack and transfers zero-shot to CFD and Crack500 while using fewer parameters and less…","keywords":["crack segmentation","EdgeSAM","Segment Anything Model","zero-shot generalization","gated attention mechanism","feature fusion","encoder-decoder architecture","pavement crack detection"],"falsifier":"Run the same DeepCrack training at 512x512 with a fine-tuned SAM-Adapter or fine-tuned EdgeSAM and evaluate zero-shot on CFD; if that baseline reaches or exceeds FlexiCrackNet's reported F1 of 54.48, the claim that frozen encoder features are the better transfer path would be contradicted.","tokens_in":15212,"feed_emoji":"🛣️","tokens_out":9756,"duration_ms":87227,"temperature":0.7,"pith_summary":"FlexiCrackNet proposes to reuse the frozen encoder of EdgeSAM, a lightweight CNN-based version of the Segment Anything Model, as a generic feature extractor for crack segmentation, then fuse those general features into a U-Net that is trained on crack images. The authors report that the resulting pipeline beats twelve comparison methods on the DeepCrack test set and in zero-shot transfer to the CFD and Crack500 datasets, while using fewer parameters and shorter inference time than the previous best crack-specific model. If the results are right, specialized segmentation models can inherit general visual knowledge from large pre-trained encoders without fine-tuning the foundation model or accepting its fixed 1024x1024 input resolution. That matters because foundation-model adaptation for small, specialized datasets is usually expensive and memory-hungry.","feed_headline":"Frozen SAM features beat crack-segmentation state of the art","feed_subtitle":"A gated-attention fusion of a frozen encoder with a U-Net also transfers to unseen road datasets.","key_machinery":"The load-bearing mechanism is IGAM, the information-interaction gated attention mechanism. At each of five encoder stages, the crack-specific feature map from the U-Net branch and the corresponding generic feature map from the frozen EdgeSAM encoder are resized and channel-aligned by a scaling module, concatenated, and passed through an information interaction module that emits two attention masks: one that selects meaningful regions in the general features and one that enhances relevant details in the crack features. The masked maps are added together and combined with a residual connection that preserves the original crack-specific features. IGAM's job is to let the decoder use the generic prior without letting non-crack textures and semantics from the general encoder contaminate the segmentation.","core_discovery":"The paper's central claim is that task-agnostic features extracted by EdgeSAM's frozen CNN encoder, never fine-tuned on cracks, are a better prior for crack segmentation than training an encoder from scratch or adapting SAM at its fixed resolution. On the DeepCrack test set, FlexiCrackNet reports F1 82.88, IoU 71.33, and Dice 82.25, surpassing CrackMamba by 1.47, 2.56, and 2.00 points. On zero-shot evaluation over the CFD dataset it reports F1 54.48, IoU 39.15, and Dice 53.83, surpassing CrackMamba by 4.50, 4.75, and 4.58 points, and on zero-shot Crack500 validation it reports F1 56.85 versus CMTFNet's 48.76. The ablation attributes the gain to the IGAM fusion: replacing IGAM with plain concatenation drops F1 from 82.88 to 78.78 on DeepCrack and from 54.48 to 35.99 on CFD, while the bare U-Net baseline sits at 79.64 and 37.48. The pipeline also reports 25.63M parameters and 22.09 ms inference time, compared with CrackMamba's 75.95M parameters and 84.73 ms.","pith_inferences":["Not tested in the paper: a direct comparison against an actual fine-tuned SAM baseline such as SAM-Adapter. If frozen features match or beat fine-tuning under the same schedule, the no-fine-tuning transfer route would be established; without that comparison, the advantage over the pre-training-plus-fine-tuning paradigm is inferred rather than demonstrated.","Not tested in the paper: input sizes other than 512x512. The claimed arbitrary-resolution flexibility follows from removing EdgeSAM's decoder, but all experiments resize inputs to 512x512, so a resolution sweep would be needed to confirm the design claim.","The fusion mechanism is domain-agnostic, so the same frozen-encoder-plus-gated-fusion recipe could plausibly transfer to other fine-grained segmentation tasks with small datasets, such as medical or industrial defect segmentation; the paper gestures at this broader applicability in its conclusion but does not test it."],"forward_implications":["On the three benchmark settings the paper tests, DeepCrack supervised plus CFD and Crack500 zero-shot, FlexiCrackNet reports the highest F1, IoU, and Dice among the twelve compared methods.","Because only EdgeSAM's CNN encoder is used and its transformer decoder is dropped, the 1024x1024 input constraint of EdgeSAM no longer applies to the pipeline, which is what the paper calls flexible input resolution.","The model is smaller and faster than the leading crack-specific baseline, with 25.63M parameters and 22.09 ms inference versus 75.95M and 84.73 ms for CrackMamba, while reporting higher accuracy on both tested domains.","The IGAM ablation indicates that how generic features are fused matters: naive concatenation can hurt performance, while the gated attention recovers and exceeds the U-Net baseline."],"supporting_citations":[{"why":"Supplies the frozen EdgeSAM CNN encoder whose multi-level features are the generic prior that the whole pipeline is built on.","marker":"[24]"},{"why":"SAM is the foundation model whose large-scale pretraining and segmentation knowledge EdgeSAM distills, grounding the claim that general visual priors transfer.","marker":"[18]"},{"why":"U-Net is the trainable encoder-decoder baseline whose skip-connection architecture FlexiCrackNet adopts and fuses the prior features into.","marker":"[29]"},{"why":"SAM-Adapter exemplifies the pre-training-plus-fine-tuning paradigm that FlexiCrackNet argues is constrained by fixed resolution and memory cost.","marker":"[20]"},{"why":"This is the closest existing approach, fine-tuning SAM for road crack segmentation, which the paper positions its frozen-feature pipeline against conceptually.","marker":"[23]"},{"why":"CrackMamba is the strongest competing crack-specific method and supplies the numerical baseline that FlexiCrackNet claims to beat on all metrics.","marker":"[48]"},{"why":"The DeepCrack dataset provides the training and test images and annotations used for the supervised comparison.","marker":"[38]"},{"why":"The CFD dataset is the zero-shot cross-domain target used to measure generalization to thin asphalt cracks.","marker":"[39]"},{"why":"The Crack500 dataset is the second zero-shot target used to measure generalization to thick cracks on textured pavement.","marker":"[40]"}],"fun_headline_variants":["Frozen SAM beats tuned nets at crack segmentation","Gated attention turns frozen SAM into crack specialist","Zero-shot crack segmentation with frozen SAM encoder","FlexiCrackNet: frozen SAM features win on cracks","SAM frozen, crack accuracy fresh: FlexiCrackNet"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the frozen EdgeSAM encoder's generic features, extracted without any crack-specific fine-tuning, are a better and more transferable source of visual prior for cracks than features from models trained or fine-tuned on crack data; the paper's evidence for this is its own reported experiments, and it does not include a direct fine-tuned-SAM comparison to test it.","fun_headline_variants_meta":{"raw":{"variants":["Frozen SAM beats tuned nets at crack segmentation","Gated attention turns frozen SAM into crack specialist","Zero-shot crack segmentation with frozen SAM encoder","FlexiCrackNet: frozen SAM features win on cracks","SAM frozen, crack accuracy fresh: FlexiCrackNet"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000298,"raw_usage":{"total_tokens":1794,"prompt_tokens":1085,"completion_tokens":709,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":633}},"tokens_in":701,"tokens_out":709,"duration_ms":8301,"temperature":1.0,"reasoning_tokens":633,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T22:11:19.861914+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same DeepCrack training at 512x512 with a fine-tuned SAM-Adapter or fine-tuned EdgeSAM and evaluate zero-shot on CFD; if that baseline reaches or exceeds FlexiCrackNet's reported F1 of 54.48, the claim that frozen encoder features are the better transfer path would be contradicted.","supporting_citations":[{"cited_title":"Segment anything,","cited_arxiv_id":null,"evidence_quote":"SAM is the foundation model whose large-scale pretraining and segmentation knowledge EdgeSAM distills, grounding the claim that general visual priors transfer."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"U-Net is the trainable encoder-decoder baseline whose skip-connection architecture FlexiCrackNet adopts and fuses the prior features into."},{"cited_title":"Sam-adapter: Adapting segment anything in underperformed scenes,","cited_arxiv_id":null,"evidence_quote":"SAM-Adapter exemplifies the pre-training-plus-fine-tuning paradigm that FlexiCrackNet argues is constrained by fixed resolution and memory cost."},{"cited_title":"Teaching segment- anything-model domain-specific knowledge for road crack segmentation from on-board cameras,","cited_arxiv_id":null,"evidence_quote":"This is the closest existing approach, fine-tuning SAM for road crack segmentation, which the paper positions its frozen-feature pipeline against conceptually."},{"cited_title":"Topology-aware mamba for crack segmentation in structures,","cited_arxiv_id":null,"evidence_quote":"CrackMamba is the strongest competing crack-specific method and supplies the numerical baseline that FlexiCrackNet claims to beat on all metrics."},{"cited_title":"Deepcrack: A deep hierarchical feature learning architecture for crack segmentation,","cited_arxiv_id":null,"evidence_quote":"The DeepCrack dataset provides the training and test images and annotations used for the supervised comparison."},{"cited_title":"Automatic road crack detection using random structured forests,","cited_arxiv_id":null,"evidence_quote":"The CFD dataset is the zero-shot cross-domain target used to measure generalization to thin asphalt cracks."},{"cited_title":"Feature pyramid and hierarchical boosting network for pavement crack detection,","cited_arxiv_id":null,"evidence_quote":"The Crack500 dataset is the second zero-shot target used to measure generalization to thick cracks on textured pavement."}],"review_version":1}