{"id":"b1b7f0fa-3f03-4c11-a70a-9899e1136391","arxiv_id":"2505.21152","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"RoBiS, combining Swin-Cropping, augmentation, adaptive thresholding, and SAM refinement, reports SegF1 of 51.00% and 46.52% on MVTec AD 2 private test sets.","lead":"This paper introduces RoBiS, a pipeline for finding and outlining defects in high-resolution industrial images. It combines overlapping image cropping, photometric augmentation, and two thresholding and refinement stages, and reports sizable gains on the MVTec AD 2 benchmark.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Swin-Cropping's stated small-anomaly preservation is undermined by the 1024-to-518 resize: a 5-pixel anomaly shrinks to ~1.3 px², below the encoder's discriminative resolution, so the central mechanism for tiny-defect gains is unverified.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing point: preprocessing at 1024x1024 followed by resizing to 518x518 must preserve the smallest anomalies for the reported SegF1 gains to reflect the paper's claimed mechanism. My stress-test adds a quantitative reason to worry: a 5-pixel anomaly is reduced to roughly 1.3 px², below the patch size of the ViT-B/14 encoder and below the SAM feature resolution the authors themselves cite as a limitation. The external official-platform evaluation makes the headline numbers credible as reported, so I do not argue for rejection. But because the core contribution is explicitly framed as solving small-anomaly localization in high-resolution images, the unresolved fate of 5-20 pixel defects is the single most load-bearing uncertainty. A synthetic defect-size sweep through the actual preprocessing and inference pipeline is the one concrete check that would settle whether the mechanism works as claimed. The reader's conditional verdict already captures this uncertainty; my analysis does not move it.","tokens_in":8036,"tokens_out":5075,"duration_ms":57726,"concrete_test":"Take 100 normal MVTec AD 2 test crops and inject synthetic defects of area 5, 10, 20, 50, and 100 pixels at realistic contrast levels; run them through the exact 1024-to-518 bilinear resize, INP-Former, and the full binarization plus SAM pipeline, then measure recall and per-defect IoU against original-resolution ground truth, binned by defect area. If defects of 20 pixels or smaller are detected at near-zero recall, the 'preserving small anomalies' rationale in Section 2.1 fails and the contribution must be restated as medium/large-defect segmentation rather than small-anomaly preservation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline SegF1 numbers are externally computed by the official challenge platform, so I do not dispute them as reported. The load-bearing point is the causal claim that Swin-Cropping 'preserves the information of small anomalies'. Section 2.1 states that over 20% of anomalies occupy fewer than 281 pixels and that the smallest anomaly contains only 5 pixels; Section 2.2.2 then resizes every 1024x1024 crop to 518x518. A 5-pixel anomaly becomes roughly 1.3 px² after bilinear downsampling, and even a 20-pixel anomaly becomes about 5 px². With a ViT-B/14 encoder (patch size 14) and the authors' own admission in Section 3.2 that SAM's 64x64 feature resolution 'prevents more precise segmentation of tiny anomalies', such defects are at or below the system's discriminative limit. The category-level numbers are consistent with this failure mode: Can (1.86/0.84) and Wall Plugs (14.38/3.40) are far below the mean, while texture categories such as Fabric and Rice dominate the average. Thus the proposed mechanism for high-resolution small-defect preservation is not established; the reported gains may be carried primarily by medium and large defects. This does not invalidate the official SegF1 numbers, but it leaves the central design motivation unresolved and makes the method's claimed advantage for tiny industrial defects unproven.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents RoBiS, a pipeline for binary defect segmentation on the MVTec AD 2 benchmark. It combines three components: (1) Swin-Cropping, a sliding-window preprocessing strategy that crops 1024x1024 patches with 10% overlap from high-resolution images; (2) an INP-Former baseline trained with Gaussian noise and exposure-based photometric augmentation; and (3) a post-processing stage that merges sub-image anomaly maps, binarizes them with a combination of mean+3std and the authors' prior MEBin module, and refines the result with SAM. The authors report SegF1 improvements from 21.8% to 51.00% on Test_private and from 16.7% to 46.52% on Test_private_mixed, with all private-test metrics computed by the official evaluation platform. Ablations on the public TESTpub set are provided for the cropping, augmentation, MEBin, and SAM-Finer components.","tokens_in":8326,"tokens_out":4853,"duration_ms":57705,"significance":"If the reported private-test numbers are accurate, the results represent a substantial improvement over the published baselines on a challenging industrial benchmark, and the pipeline is simple enough to be useful in practice. The paper has concrete strengths: the headline SegF1 values are externally computed by the official challenge platform rather than self-reported, code is promised at a public repository, and the framework is decomposed into clearly described modules. However, the causal explanation for the improvement is not fully supported by the experiments as presented: the claim that Swin-Cropping preserves small-anomaly information is undermined by the subsequent resize step, and the ablation tables lack the no-module baseline needed to attribute gains to individual components. The paper would be strengthened by additional experiments that directly test the small-anomaly mechanism and by completing the ablation design.","major_comments":[{"comment":"The central design motivation for Swin-Cropping is that it 'preserves the information of small anomalies' by cropping high-resolution images, but the manuscript does not establish that this mechanism actually survives the subsequent resize. Section 2.1 reports that over 20% of anomalies occupy fewer than 281 pixels and that the smallest anomaly contains only 5 pixels; Section 2.2.2 then resizes every 1024x1024 crop to 518x518. A 5-pixel anomaly is reduced to roughly 1.3 square pixels after bilinear downsampling, and the authors themselves note in Section 3.2 that SAM's 64x64 feature resolution 'prevents more precise segmentation of tiny anomalies.' The category-level results in Table 1 are consistent with a failure on very small defects: Can (1.86/0.84) and Wall Plugs (14.38/3.40) are far below the mean, while texture categories such as Fabric and Rice dominate the average. To support the claimed mechanism, please provide a quantitative analysis of performance as a function of anomaly size, or an experiment comparing cropping with and without the 518x518 resize on categories with tiny defects, or a synthetic downsampling study that shows small anomalies are not destroyed before detection.","section":"Sec. 2.1, Sec. 2.2.2, Table 1"},{"comment":"The ablation tables do not include a baseline row with neither module active, so the claimed contributions of individual components cannot be computed. In Table 3, the text states that Swin-Cropping gives a 3.7% F1-max improvement and data augmentation gives an additional 1.9%, but the table only shows three configurations (Swin only, augmentation only, and both); without a 'neither' row, the 3.7% is a marginal gain over augmentation alone and the 1.9% is a marginal gain over Swin-Cropping alone, not an isolation of either module. The same issue applies to Table 4, where the claimed 0.4% gain for MEBin and 12.5% gain for SAM-Finer require a no-module baseline that is absent. Additionally, these ablations are reported only on TESTpub, while the headline claims are on the private test sets; please add the missing baseline row and, where possible, report the corresponding ablations on the private sets or justify why TESTpub is representative.","section":"Tables 3 and 4"},{"comment":"The choice of SAM variant is test-set dependent: the authors state 'we use SAM-Huge architecture for TESTpriv and SAM-Base for TESTpriv,mix' without giving a rationale or an ablation. This is a free design parameter applied differently to the two private test sets, and if the choice was made using feedback from the official evaluation platform, it affects the interpretation of the reported numbers as a fixed unsupervised pipeline. Please provide a justification for this choice, report public-set validation for both variants, or state explicitly how the variant was selected.","section":"Sec. 2.3"}],"minor_comments":[{"comment":"The sentence beginning 'While our method successfully identifies both structural and logical anomalies, architectural constraints inherent to anomaly detection models, particularly the need for successive downsampling to achieve sufficient receptive fields.' is grammatically incomplete and should be finished.","section":"Sec. 3.2"},{"comment":"The full text says 'Code is available at github link' instead of an actual URL; please replace this placeholder with the complete repository address.","section":"Abstract and Sec. 5"},{"comment":"The term 'multi-class approaches' is confusing because the method trains one model per product category; please clarify that 'multi-class' refers to multiple sub-image appearances within a category rather than multiple semantic classes.","section":"Sec. 2.2.1"},{"comment":"The notation for the private test sets is inconsistent: the abstract uses Test_private and Test_private_mixed, while the body and tables use TESTpriv and TESTpriv,mix. Please unify the notation.","section":"Sec. 2.4 and Table 1"},{"comment":"The reference title 'The mvtec ad 2 dataset: Advanced scenarios for unsupervised anomaly detection' should be capitalized consistently as 'The MVTec AD 2 Dataset: Advanced Scenarios for Unsupervised Anomaly Detection'.","section":"Reference [12]"}],"recommendation":"major_revision","confidential_remarks":"The paper reads as a short challenge-report-style manuscript rather than a full journal article. The two main obstacles for publication are the missing no-module baselines in the ablations and the unverified small-anomaly preservation mechanism. The external official-platform numbers give the headline results credibility, but the authors need to demonstrate that the mechanism they claim is actually what drives the gains, and that the per-test-set SAM variant is not a form of post-hoc tuning."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a benchmark entry that stacks known parts and wins big on the external private test scores. The SegF1 improvements are computed by the official platform, so I'm inclined to trust them as numbers. What is genuinely new is the full pipeline—Swin-Cropping, augmentation, adaptive binarization with MEBin, SAM refinement—and its evaluation on MVTec AD 2. The paper does a decent job documenting the components and giving category-level tables, and the code is public.\n\nThe soft spots are real but not fatal. The stress-test concern about small anomalies is on point: the paper claims Swin-Cropping preserves tiny defects, but then resizes 1024x1024 crops to 518x518, so a 5-pixel anomaly becomes roughly 1.3 px². The paper itself admits SAM's 64x64 feature resolution limits tiny-anomaly segmentation. So the mechanism for small-anomaly preservation is not established, and the category results (Can, Wall Plugs) indeed lag far behind the mean. That is a missing analysis, not a reason to reject the reported numbers—the gains may mostly come from medium and large defects. The ablations on the public set lack a clean no-module baseline, and the choice of SAM-Huge for TESTpriv versus SAM-Base for TESTpriv,mix is unexplained. There is also an unfinished sentence in Sec 3.2 and some ambiguous tables.\n\nOverall, this is a solid engineering contribution for the anomaly segmentation community. It is not a scientific breakthrough—each component exists in prior work—but the integration works and the external evaluation gives it grounding. If I were an editor, I would send it to review with a request for better ablations and an honest discussion of what the method cannot do for tiny defects. It is not a desk reject.","headline":"A credible challenge entry that lifts SegF1 on MVTec AD 2 with known components, but the small-anomaly story is weaker than advertised.","tokens_in":8913,"tokens_out":1911,"would_cite":false,"duration_ms":19966,"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":"RoBiS reports that a preprocessing-plus-thresholding pipeline lifts industrial defect segmentation F1 from 21.8% to 51.00% on the private MVTec AD 2 test set, and from 16.7% to 46.52% on the mixed-lighting private test set.","keywords":["unsupervised anomaly detection","industrial defect segmentation","high-resolution image preprocessing","sliding-window cropping","adaptive binarization","MVTec AD 2","lighting augmentation","SAM refinement"],"falsifier":"Measure SegF1 on MVTec AD 2 TESTpriv with Swin-Cropping disabled (full high-resolution images resized directly to 518x518) while keeping the same augmentation and binarization; if small-defect categories such as Can and Wall Plugs do not lose most of their already-low scores, the reported gains cannot be attributed to the cropping step. The official per-category scoreboard would reveal whether the 14.38% and 1.86% SegF1 on Wall Plugs and Can are actually driven by the surviving small defects.","tokens_in":7781,"feed_emoji":"🔍","tokens_out":6612,"duration_ms":58360,"temperature":0.7,"pith_summary":"This paper argues that the steep performance drop of unsupervised anomaly detection on the MVTec AD 2 industrial benchmark can be reversed by treating the full pipeline as three separate robustness problems: keeping small defects visible in high-resolution images, hardening the detector to lighting and noise shifts, and turning continuous anomaly maps into accurate binary masks. It reports that its RoBiS pipeline raises segmentation F1 on the private test set from 21.8% to 51.00% and on the mixed-lighting private test set from 16.7% to 46.52%, gains of 29.2 and 29.82 percentage points over the compared baselines. If the numbers hold, the practical message is that preprocessing and post-processing choices, not just the anomaly-detection model, are the main remaining bottleneck for real industrial deployment.","feed_headline":"Defect segmentation F1 jumps from 21.8% to 51.00%","feed_subtitle":"Overlapping crops, lighting augmentation, and adaptive binarization drive the gain on private MVTec AD 2 tests.","key_machinery":"The load-bearing mechanism is the combination of Swin-Cropping and the joint binarization cascade. Swin-Cropping is an overlapping sliding-window preprocessing step (1024x1024 windows, 10% overlap, zero-padded boundaries) that preserves small and elongated anomalies while avoiding boundary-split artifacts; it is what lets a detector trained on 518x518 sub-images see defects that occupy as few as five pixels in the original high-resolution frame. The binarization cascade first thresholds each merged anomaly map by the per-pixel mean plus three standard deviations, takes the logical OR with MEBin's adaptively thresholded mask, then refines the result with SAM prompted by bounding boxes of the coarse regions. This cascade is what converts continuous anomaly scores into the final binary masks that the SegF1 metric is computed on.","core_discovery":"On the paper's own terms, RoBiS establishes that the central difficulty in high-resolution industrial anomaly segmentation is not the detector alone but the joint handling of scale, illumination, and thresholding. Swin-Cropping cuts each high-resolution image into overlapping 1024x1024 windows with 10% overlap, so that anomalies as small as a handful of pixels survive into the sub-images; those sub-images are resized to 518x518 and fed to INP-Former, a multi-class anomaly-detection baseline initialized with DINOv2-R features. During training, half the samples receive Gaussian noise and exposure shifts to simulate the lighting variations of the test sets. At inference, overlapping anomaly maps are merged, then binarized by the union of a mean-plus-three-standard-deviation threshold and the authors' MEBin adaptive threshold, and finally refined by prompting SAM with the bounding boxes of coarse abnormal regions. The paper reports SegF1 of 51.00% on TESTpriv and 46.52% on TESTpriv,mix, which it presents as a 29.2-percentage-point and a 29.82-percentage-point gain over the best compared methods.","pith_inferences":["A testable extension is to ablate the input resolution: because the pipeline resizes 1024x1024 crops down to 518x518, a detector that operates natively at 1024x1024 might recover even more of the small-anomaly signal, making the reported gains a lower bound on what preprocessing alone can do.","The same preprocessing-plus-adaptive-binarization recipe could transfer to other high-resolution anomaly benchmarks beyond MVTec AD 2, and the paper's segmentation gains suggest that thresholding is undervalued relative to detector architecture in the field.","Because SAM-Finer contributes the largest single post-processing gain (12.5 points on TESTpub), the paper implies that foundation-model refinement is a low-cost upgrade for any continuous anomaly map; attaching SAM-Finer to a different baseline's anomaly maps without retraining would test this directly."],"forward_implications":["Overlapping window cropping alone yields a 3.7-percentage-point F1-max improvement on TESTpub by preserving small-anomaly information in sub-images.","Adding Gaussian noise and exposure-shift augmentation to 50% of training samples contributes another 1.9 percentage points on TESTpub.","MEBin's adaptive threshold adds 0.4 percentage points of SegF1 on TESTpub by catching subtle anomalies that the simple mean-plus-three-standard-deviations threshold misses.","SAM-Finer contributes 12.5 percentage points of SegF1 on TESTpub, mainly by filling internal false negatives in fabric-class defects.","The full pipeline reaches 51.00% SegF1 on TESTpriv and 46.52% on TESTpriv,mix, with balanced ClassF1 and AucPro0.05 scores, and the code is released."],"supporting_citations":[{"why":"Supplies the MVTec AD 2 dataset, its three test splits, and the official SegF1 metric the reported gains are measured against.","marker":"[12]"},{"why":"The INP-Former baseline that produces the per-sub-image anomaly maps which RoBiS merges and binarizes.","marker":"[17]"},{"why":"Introduces the MEBin adaptive-binarization method, which the pipeline OR-combines with the mean-plus-three-standard-deviations threshold.","marker":"[13]"},{"why":"The Segment Anything model used in SAM-Finer to refine coarse binary masks from bounding-box prompts.","marker":"[14]"},{"why":"Provides the DINOv2-R pretrained weights for the ViT-B-14 encoder of the baseline model.","marker":"[6]"},{"why":"One of the Table 1 baselines whose TESTpriv/TESTpriv,mix SegF1 scores form the comparison for the claimed gain.","marker":"[20]"},{"why":"The MSFlow baseline whose 21.8% TESTpriv SegF1 is the reference starting point for the 29.2-percentage-point improvement claim.","marker":"[29]"},{"why":"Reverse-distillation baseline used in the Table 1 comparison on both private test sets.","marker":"[7]"}],"fun_headline_variants":["RoBiS lifts defect F1 from 21.8% to 51.0% on private MVTec AD 2","Overlapping crops and adaptive thresholding boost SegF1 to 51.0%","Industrial AD SegF1 gains 29 points with RoBiS's three-stage pipeline","From 21.8 to 51.0: RoBiS redefines high-res defect segmentation","RoBiS: 29.2-point SegF1 gain on high-res industrial images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes that the cropping-and-resizing step keeps defects as small as a few pixels detectable, meaning no tiny anomaly is destroyed or blurred below the detector's resolution before the model ever sees it.","fun_headline_variants_meta":{"raw":{"variants":["RoBiS lifts defect F1 from 21.8% to 51.0% on private MVTec AD 2","Overlapping crops and adaptive thresholding boost SegF1 to 51.0%","Industrial AD SegF1 gains 29 points with RoBiS's three-stage pipeline","From 21.8 to 51.0: RoBiS redefines high-res defect segmentation","RoBiS: 29.2-point SegF1 gain on high-res industrial images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000794,"raw_usage":{"total_tokens":3545,"prompt_tokens":1045,"completion_tokens":2500,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":2374}},"tokens_in":661,"tokens_out":2500,"duration_ms":17007,"temperature":1.0,"reasoning_tokens":2374,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T13:33:45.082225+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure SegF1 on MVTec AD 2 TESTpriv with Swin-Cropping disabled (full high-resolution images resized directly to 518x518) while keeping the same augmentation and binarization; if small-defect categories such as Can and Wall Plugs do not lose most of their already-low scores, the reported gains cannot be attributed to the cropping step. The official per-category scoreboard would reveal whether the 14.38% and 1.86% SegF1 on Wall Plugs and Can are actually driven by the surviving small defects.","supporting_citations":[{"cited_title":"Explor- ing intrinsic normal prototypes within a single image for universal anomaly detection","cited_arxiv_id":null,"evidence_quote":"The INP-Former baseline that produces the per-sub-image anomaly maps which RoBiS merges and binarizes."},{"cited_title":"Anomalyncd: Towards novel anomaly class discovery in industrial scenarios","cited_arxiv_id":null,"evidence_quote":"Introduces the MEBin adaptive-binarization method, which the pipeline OR-combines with the mean-plus-three-standard-deviations threshold."},{"cited_title":"Segment any- thing","cited_arxiv_id":null,"evidence_quote":"The Segment Anything model used in SAM-Finer to refine coarse binary masks from bounding-box prompts."},{"cited_title":"Vision transformers need registers","cited_arxiv_id":null,"evidence_quote":"Provides the DINOv2-R pretrained weights for the ViT-B-14 encoder of the baseline model."},{"cited_title":"Towards total recall in industrial anomaly detection","cited_arxiv_id":null,"evidence_quote":"One of the Table 1 baselines whose TESTpriv/TESTpriv,mix SegF1 scores form the comparison for the claimed gain."},{"cited_title":"Msflow: Multiscale flow-based framework for unsupervised anomaly detection","cited_arxiv_id":null,"evidence_quote":"The MSFlow baseline whose 21.8% TESTpriv SegF1 is the reference starting point for the 29.2-percentage-point improvement claim."},{"cited_title":"Anomaly detection via reverse distillation from one-class embedding","cited_arxiv_id":null,"evidence_quote":"Reverse-distillation baseline used in the Table 1 comparison on both private test sets."}],"review_version":1}