{"id":"cad96d38-c046-4a82-8bbf-004df1476a5b","arxiv_id":"2501.12295","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A multi-level transformer-CNN reconstruction model, UniAS, reports new state-of-the-art pixel-level anomaly segmentation scores on MVTec-AD and VisA, but its Dice scores rely on oracle thresholds chosen from test labels.","lead":"This paper introduces UniAS, a unified neural network that segments image defects pixel-by-pixel by reconstructing normal features at multiple scales. It reports large gains on standard industrial anomaly benchmarks over previous one-for-all detectors, though part of the gain comes from a threshold chosen using test labels.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The DSC half of the SOTA claim depends on test-label oracle thresholds and a non-standard empty-GT rule, so the 'significant surpassing' margin is not deployment-realistic; pAP alone leaves only a slim VisA advantage.","rationale":"The reader's weakest-assumption identification is correct and is the most load-bearing point: the DSC half of the headline result is not a deployable number because the segmentation threshold is chosen on test ground truth (Sec. 4.1; Supp. A.2), and the special empty-GT rule further distorts the metric. I considered whether the SAR query (Sec. 3.4) could leak test-time anomaly information into the reconstruction and artificially inflate scores, but the mechanism is a learned reweighting and the paper's ablations show it helps; this is not a clear correctness flaw. I also considered the small pAP margin on VisA, but pAP is threshold-free and still places UniAS first, so it supports a competitive claim even if not a 'large margin' on that dataset. The most serious issue remains the DSC evaluation protocol, which directly affects one of the two headline metrics and the strength of the 'significantly surpassing previous methods' statement. Since all baselines are re-evaluated under the same oracle protocol, the relative ranking may persist, but the absolute DSC values and the claimed large margins are not what an unsupervised system would achieve. The paper provides code and detailed ablations, which are positive evidence, but the central SOTA claim should be conditioned on a label-free threshold evaluation. This does not move the verdict beyond the reader's CONDITIONAL recommendation; it reinforces it.","tokens_in":18161,"tokens_out":10053,"duration_ms":107051,"concrete_test":"Recompute the DSC tables (Table 3 and Supp. Tab. 1-2) using a label-free threshold: (i) set a per-category threshold from the score distribution on normal training images (e.g., the 99th percentile of anomaly scores on normal pixels); or (ii) apply per-image Otsu thresholding to each anomaly map; or (iii) hold out a labeled validation subset of test images and select the threshold there. Also recompute DSC excluding normal images or using standard handling of empty GT (e.g., count DICE as 1 only if both masks are empty, otherwise exclude). If UniAS's DSC margin over HVQ-Trans/RD4AD drops below the 'large margin' claimed in Sec. 4.6, the DSC SOTA claim is protocol-dependent. Report pAP with standard errors across at least three seeds to verify the VisA +1.83 margin is stable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline quantitative claim (Abstract; Sec. 4.6) rests on two metrics. pAP is threshold-free and UniAS does rank first, but on VisA the margin over RD4AD is only 1.83 points (40.06 vs 38.23). The DSC numbers (59.33/32.50) are computed under a protocol that selects, per category, the threshold maximizing (precision+recall) on the test-set PR curve (Sec. 4.1; Supp. A.2). That requires ground-truth masks, so these Dice values are oracle-threshold results, not what a deployment without labels would obtain. Additionally, Supp. A.2 defines DSC for normal images (empty GT) as 1 if the predicted mask is empty and 0 if any pixel is predicted; this makes DSC heavily reward predicting 'nothing' on the roughly 50% normal test images and can inflate or deflate scores depending on threshold placement. Because all baselines are re-evaluated under the same protocol, the relative ordering may survive, but the 'large margin' (e.g., +26.70 DSC over HVQ-Trans on MVTec, Sec. 4.6) is a property of an oracle-thresholded, non-standard metric rather than of deployable anomaly segmentation. The central 'surpassing previous methods significantly' claim is therefore only partially supported as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes UniAS, a unified one-for-all unsupervised anomaly segmentation model that reconstructs multi-level features through a hybrid Transformer and multi-granularity gated CNN (MGG-CNN) pipeline, with a sample-aware reweighted query. The authors argue that AUROC is an inflated metric under imbalanced anomaly settings and advocate pAP and DSC as segmentation-oriented metrics. On MVTec-AD and VisA they report 65.12/59.33 and 40.06/32.50 pAP/DSC, respectively, claiming state-of-the-art anomaly segmentation with significant improvements over prior methods.","tokens_in":18503,"tokens_out":6610,"duration_ms":67212,"significance":"If the results hold, UniAS offers a practical single-model, per-class-fine-tuning-free anomaly segmentation approach and contributes a useful critique of AUROC for imbalanced pixel-level evaluation. The threshold-free pAP ranking is credible, the multi-level design is supported by the ablation study in Sec. 4.7, and the code is shared, which strengthens reproducibility. The principal caveat is that the DSC half of the headline result depends on an oracle-threshold evaluation protocol that is not deployment-realistic; this affects the strength of the central SOTA claim as stated.","major_comments":[{"comment":"The reported DSC values are obtained by choosing, for each category, the binarization threshold that maximizes precision plus recall on the test-set PR curve, using ground-truth masks. This is an oracle-threshold quantity, not a label-free prediction, and the large margins quoted in the Abstract and Sec. 4.6 (e.g., +26.70 DSC over HVQ-Trans on MVTec) are properties of that protocol rather than of a deployable anomaly-segmentation pipeline. The pAP metric is threshold-free and is not affected, but the DSC half of the headline claim should be re-reported under a threshold-free or fixed-threshold protocol, or clearly labeled as oracle-threshold DSC throughout.","section":"Sec. 4.1 / Supp. A.2"},{"comment":"The empty-ground-truth DSC rule for normal images (DSC = 1 if the predicted mask is empty, 0 if any pixel is predicted) is non-standard and, because a large fraction of test images are normal, it can dominate category-mean DSC and can either inflate or depress scores depending on where the oracle threshold lands. Please report DSC on anomalous test images only, and/or with a standard convention for empty ground truth, and quantify how this empty-GT rule affects the reported means.","section":"Supp. A.2"},{"comment":"The claim that UniAS 'surpasses previous methods significantly' is strongly supported for MVTec pAP (65.12 vs 59.20 for DeSTSeg), but on VisA the pAP margin over the next-best method is 1.83 points (40.06 vs 38.23 for RD4AD), and no variance or repeated-run information is provided. The strength of the 'significant' claim should be qualified, or supported with standard errors or confidence intervals.","section":"Abstract / Sec. 4.6"}],"minor_comments":[{"comment":"The supplementary material contains unresolved citation placeholders ('[?]') and no corresponding reference list; these must be completed before publication.","section":"Supp. A.1 / A.2 / Supp. Tables"},{"comment":"The label '25’W ACV' for the proposed method should read 'WACV 2025' (or a consistent citation string for the venue).","section":"Tables 2 and 3"},{"comment":"There is a typo: 'datatset' should be 'dataset'.","section":"Supp. B.3.2"},{"comment":"pAP and DSC are described but not defined by equations; since the paper argues for their use as primary metrics, formal definitions or explicit citations would improve clarity.","section":"Sec. 4.1"},{"comment":"Several per-class DSC values are near zero (e.g., VisA Macaroni2 0.07, Pcb2 0.69); a brief comment on threshold degeneracy for low-anomaly-rate classes would help readers interpret the stability of the DSC protocol.","section":"Supp. Tables 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The core pAP finding is credible and the architecture ablation supports the multi-level hybrid design, so I do not see grounds for rejection. However, the DSC-based SOTA claim is currently tied to an oracle-threshold protocol and a non-standard empty-GT rule, and the VisA pAP margin is modest. The revision should reframe or supplement the DSC reporting, and the supplementary references must be completed. The paper fits the scope of the venue but needs this load-bearing evaluation issue addressed before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent, useful paper, but read the DSC numbers with a very skeptical eye. The pAP result looks real; the DSC half of the SOTA claim is an oracle-threshold artifact.\n\nWhat's genuinely new: the UniAS architecture—Mask2Former-style multi-level decoding with a gated multi-granularity CNN and a sample-aware reweighted query—is a sensible combination that clearly beats prior one-for-all feature reconstruction methods on threshold-free pixel AP. The ablation study actually supports the design: stacking more feature levels monotonically improves pAP, and the MGG-CNN/gate components each add a clear increment. The paper also makes a fair case that AUROC saturates and hides false positives, and it correctly points to pAP and DSC as more informative segmentation metrics. That argument isn't new—Bertoldo et al. and Rafiei et al. said it earlier—but the authors cite them and apply them, which is fine.\n\nWhere it gets soft. The headline DSC figures (59.33/32.50) are computed with per-category thresholds selected on the test-set PR curve using ground-truth masks. That's an oracle threshold; no deployment would have that. The supplement also defines DSC for normal images as 1 if the predicted mask is empty and 0 if any pixel is predicted, which is non-standard and can inflate scores on the roughly half of test images that are normal. These protocol choices are disclosed in the supplement, but the main text says 'maximizing the sum of precision and recall on the PR curve' without flagging that this uses test labels. So the 'surpassing previous methods significantly' claim rests on a metric that rewards knowing the answer. The pAP advantage is threshold-free and credible, but even there, the VisA margin over RD4AD is 1.83 points—slim.\n\nAlso, the supplement has several placeholder citation markers ('[?]'), including for CBAM and the dual transformer design. The authors need to clean that up before publication; it looks unfinished.\n\nWho should read this: anyone working on unified anomaly detection or on evaluation metrics for imbalanced segmentation. It's a solid empirical contribution, not a conceptual breakthrough. A referee can do a useful job: ask for label-free thresholds, per-run variance, and proper citations. It deserves peer review, and with those fixes it could be a solid WACV-level paper.","headline":"Solid one-for-all anomaly segmentation with a credible pAP win; the DSC headline is oracle-thresholded and overstated, but the architecture and analysis deserve referee time.","tokens_in":19004,"tokens_out":2316,"would_cite":true,"duration_ms":23918,"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":"UniAS claims that multi-level hybrid reconstruction with gated multi-granularity CNNs segments industrial anomalies at pixel level, reaching pAP/DSC of 65.12/59.33 on MVTec-AD and 40.06/32.50 on VisA, surpassing previous unified models.","keywords":["anomaly segmentation","unsupervised anomaly detection","one-for-all","multi-level reconstruction","MGG-CNN","pixel-wise average precision","Dice similarity coefficient","industrial defect detection"],"falsifier":"Recompute UniAS and its baselines on MVTec-AD and VisA with a single threshold derived from normal training images, such as the 95th percentile of anomaly scores on normal validation data, instead of per-class oracle thresholds, and compare pAP and DSC; if the margin over HVQ-Trans and UniAD largely disappears, the headline DSC numbers depend on oracle thresholding rather than on the model's segmentation ability.","tokens_in":17988,"feed_emoji":"🔍","tokens_out":6438,"duration_ms":61637,"temperature":0.7,"pith_summary":"Unsupervised anomaly segmentation is usually scored by AUROC, but this paper argues that AUROC is inflated when anomalies occupy only a few percent of pixels, as in the MVTec and VisA datasets. The authors propose UniAS, a single one-for-all model that reconstructs normal features at multiple resolutions and combines the reconstruction errors into a pixel-level anomaly map. The central claim is that multi-level hybrid reconstruction, adding a multi-granularity gated CNN to transformer layers, preserves the low-level detail needed to segment defect boundaries while suppressing false-positive regions. On the two benchmarks the model reports pAP/DSC of 65.12/59.33 and 40.06/32.50, surpassing prior unified and per-class methods on these segmentation metrics. A reader should care because accurate pixel-level localization is what is actually needed for monitoring and removing industrial defects, and a single model would remove the cost of per-class training.","feed_headline":"65.12 pAP: one model sharpens industrial anomaly segmentation","feed_subtitle":"Multi-level transformer-CNN reconstruction suppresses the false positives that inflate AUROC scores.","key_machinery":"The load-bearing mechanism is the multi-level hybrid decoder. At each level, features from a frozen CNN backbone are patchified and fed as keys and values to a transformer layer, whose cross-attention with a learnable query reconstructs the normal pattern globally; the query then passes through the MGG-CNN, four parallel convolutional branches gated by GELU with receptive fields of 1, 3, 5, and 7, which adds local detail at multiple granularities. A transposed convolution unpatchifies the result back to the original feature resolution, and training minimizes cosine distance plus mean squared error against the extracted features. At inference the per-level cosine distances are upsampled to image size and multiplied pixelwise, an operation that suppresses false positives because a pixel must be anomalous at every level to survive. A sample-aware reweighting module, using channel and spatial attention computed from the highest-level feature, reweights the random initial query so that one-for-all reconstruction adapts to the input sample.","core_discovery":"The paper's discovery is that prior one-for-all anomaly detectors, by downsampling features and reconstructing at a coarse scale, trade away exactly the low-level information that pixel-level segmentation needs. UniAS instead reconstructs features at each of four levels in a coarse-to-fine decoder: a transformer layer supplies global context and a multi-granularity gated CNN (MGG-CNN) refines local details at receptive fields of 1, 3, 5, and 7 pixels. During inference, the cosine distance between reconstructed and extracted features yields an anomaly map per level, and the maps are upsampled and multiplied together; the authors show these maps play complementary roles, with high levels finding the anomaly semantically and low levels tracing its contour. With this design, plus a sample-aware reweighted query that makes the shared memory query more specific to each input, UniAS reports state-of-the-art pAP/DSC of 65.12/59.33 on MVTec-AD and 40.06/32.50 on VisA, and its ablations attribute most of the gain to the multi-level pipeline and the gated multi-granularity branch.","pith_inferences":["Editorial extension: the reported DSC values are computed with per-category thresholds selected on the test set by maximizing precision and recall; a deployment-oriented evaluation would set thresholds from normal training data only, and the comparative advantage may be smaller under that protocol.","Editorial extension: the same coarse-to-fine multiplicative fusion could be tested on medical anomaly segmentation, where lesions are often small and boundary accuracy is clinically important.","Editorial extension: because the feature-combination ablation shows the lowest level helps most when added to coarser levels, the method's gain may come less from the transformer than from preserving the shallowest features; an experiment that holds the decoder fixed and varies only the set of levels would separate these factors.","Editorial extension: the Gaussian-filter preprocessing with residual concatenation is presented as a detail, but the supplementary ablation shows it changes DSC by several points; varying filter sigma could be a cheap hyperparameter to tune further."],"forward_implications":["Pixel-level AUROC should no longer be the headline metric for anomaly segmentation; pAP and DSC expose false positives that AUROC hides.","A single unified model trained only on normal images can segment defects across many industrial object categories at pixel level, removing per-class fine-tuning.","Multi-level reconstruction with multiplicative fusion is a transferable design choice for other reconstruction-based detection models.","Fine anomaly boundaries, such as the thin defects in MVTec, become segmentable when low-level features are reconstructed rather than discarded.","The SAR query suggests that sample-adaptive initialization of shared memory queries can improve cross-category reconstruction without extra per-category parameters."],"supporting_citations":[{"why":"UniAD, the unified transformer baseline whose high-AUROC but low-DSC behavior motivates the paper and that UniAS is compared against.","marker":"[45]"},{"why":"HVQ-Trans, the closest one-for-all transformer baseline that UniAS claims to surpass by a large margin on pAP and DSC.","marker":"[27]"},{"why":"MVTec-AD, the main industrial anomaly dataset and source of the pixel-level ground truths used for segmentation evaluation.","marker":"[1]"},{"why":"VisA, the second benchmark with tiny anomalies and noisy backgrounds that UniAS is evaluated on.","marker":"[54]"},{"why":"Supplies the argument that AUROC is inflated for pixel-level anomaly detection under class imbalance.","marker":"[28]"},{"why":"Supplies the cosine-distance plus MSE reconstruction loss used to train the decoder.","marker":"[32]"},{"why":"Supplies the multi-level transformer decoder structure that UniAS adapts for one-for-all reconstruction.","marker":"[8]"},{"why":"PatchCore, the memory-bank baseline whose expensive memory is contrasted with the SAR query and which is compared in pAP and DSC.","marker":"[29]"},{"why":"Supplies the channel and spatial attention used in the sample-aware reweighting of the query.","marker":"[39]"}],"fun_headline_variants":["One model fixes anomaly segmentation AUROC blindspot","UniAS: coarse-to-fine anomaly maps beat AUROC games","65.12 pAP: multi-level fusion nails pixel anomalies","Anomaly segmentation: gated CNN + transformer at four scales","Say goodbye to AUROC inflation: UniAS localizes precisely"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that comparing models on pAP and DSC with per-category thresholds tuned on the test set's ground-truth masks is the right way to measure anomaly segmentation; if those masks are not available at deployment, the reported Dice advantage may not survive.","fun_headline_variants_meta":{"raw":{"variants":["One model fixes anomaly segmentation AUROC blindspot","UniAS: coarse-to-fine anomaly maps beat AUROC games","65.12 pAP: multi-level fusion nails pixel anomalies","Anomaly segmentation: gated CNN + transformer at four scales","Say goodbye to AUROC inflation: UniAS localizes precisely"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000717,"raw_usage":{"total_tokens":3237,"prompt_tokens":978,"completion_tokens":2259,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":594,"completion_tokens_details":{"reasoning_tokens":2173}},"tokens_in":594,"tokens_out":2259,"duration_ms":14773,"temperature":1.0,"reasoning_tokens":2173,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:18:39.564113+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute UniAS and its baselines on MVTec-AD and VisA with a single threshold derived from normal training images, such as the 95th percentile of anomaly scores on normal validation data, instead of per-class oracle thresholds, and compare pAP and DSC; if the margin over HVQ-Trans and UniAD largely disappears, the headline DSC numbers depend on oracle thresholding rather than on the model's segmentation ability.","supporting_citations":[{"cited_title":"A unified model for multi-class anomaly detection","cited_arxiv_id":null,"evidence_quote":"UniAD, the unified transformer baseline whose high-AUROC but low-DSC behavior motivates the paper and that UniAS is compared against."},{"cited_title":"Hierarchical vector quantized transformer for multi-class unsupervised anomaly detection","cited_arxiv_id":null,"evidence_quote":"HVQ-Trans, the closest one-for-all transformer baseline that UniAS claims to surpass by a large margin on pAP and DSC."},{"cited_title":"Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection","cited_arxiv_id":null,"evidence_quote":"MVTec-AD, the main industrial anomaly dataset and source of the pixel-level ground truths used for segmentation evaluation."},{"cited_title":"On Pixel-level Performance Assessment in Anomaly Detection","cited_arxiv_id":"2310.16435","evidence_quote":"Supplies the argument that AUROC is inflated for pixel-level anomaly detection under class imbalance."},{"cited_title":"Multiresolution knowledge distillation for anomaly detection","cited_arxiv_id":null,"evidence_quote":"Supplies the cosine-distance plus MSE reconstruction loss used to train the decoder."},{"cited_title":"Masked-attention mask transformer for universal image segmentation","cited_arxiv_id":null,"evidence_quote":"Supplies the multi-level transformer decoder structure that UniAS adapts for one-for-all reconstruction."},{"cited_title":"Towards to- tal recall in industrial anomaly detection","cited_arxiv_id":null,"evidence_quote":"PatchCore, the memory-bank baseline whose expensive memory is contrasted with the SAR query and which is compared in pAP and DSC."}],"review_version":1}