{"id":"07f6f8cd-f01d-4abb-9660-785e245c4af9","arxiv_id":"2508.12927","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"PRADOT learns local and global prototypes via optimal transport with a fused feature-spatial cost, achieving competitive anomaly detection and localization on industrial benchmarks.","lead":"This paper presents PRADOT, an unsupervised anomaly detection method that learns local and global prototypes using optimal transport, with a cost that balances feature similarity and spatial proximity. On MVTec AD LOCO it ranks second among strong baselines for classification and localization, showing an advantage on logical anomalies.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ablation varies α jointly in training and inference, so the claimed benefit of learning local prototypes is not established.","rationale":"The reader's verdict (CONDITIONAL) is reasonable; I would keep it. My main concern differs from the reader's weakest_assumption (uniform OT marginals): the uniform marginal is a plausible but secondary design limitation. The more load-bearing issue is attribution: the paper's headline contribution is the structural constraint during prototype learning, but the reported ablation varies α in a way that cannot separate learning from inference. This is addressable with a small set of cross-configuration runs and does not invalidate the empirical results; it only determines what the results actually show. The reader also noted the main-text update equation (Eq. 5) is inconsistent with the pseudocode; that is a reproducibility concern that reinforces CONDITIONAL, but I did not make it the central attack because the pseudocode and available code likely resolve it. Overall, no change to the reader's verdict is required beyond keeping it conditional pending the separation test.","tokens_in":13803,"tokens_out":7741,"duration_ms":83134,"concrete_test":"Run a 2x2 ablation on MVTec AD LOCO logical anomalies: (a) train α=0, infer α=0; (b) train α=0, infer α=0.5; (c) train α=0.5, infer α=0; (d) train α=0.5, infer α=0.5. If AU-sPRO for logical anomalies in (b) reaches the level of (d), the improvement is an inference-side effect and the 'learning local prototypes' claim needs revision; if (c) alone matches (d), learning is the active ingredient.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that enforcing a structural constraint when learning prototypes improves detection of incoherencies. But α enters both the training cost (Eq. 1, used in the OT cost matrix Eq. 3 and in the prototype update through Eq. 4) and the inference assignment (Eq. 6). The ablation in Section 4.3 (Fig. 3b) varies α jointly and reports that higher α improves logical anomaly localization. This does not distinguish whether the gain comes from learning spatially organized prototypes or merely from applying the spatial penalty at test time. The text even describes the effect as 'imposing a structural constraint on the embedding-to-prototype assignment' (§4.3), which is an inference-side statement. If the same improvement is obtained by training with α=0 and scoring with α=0.5, the prototype learning mechanism is not responsible for the advertised advantage; the abstract's 'when learning the prototypes' claim would be unsupported. The uniform-marginal assumption (Eq. 2) identified by the reader is a limitation of the modeling choice, not the decisive point here.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PRADOT, an unsupervised anomaly detection and localization method that learns feature-space prototypes from a pre-trained encoder under a fused feature/spatial cost inspired by the fused Gromov-Wasserstein distance. Embeddings and prototypes carry spatial coordinates on a fixed grid; during training, Sinkhorn optimal transport assigns batch embeddings to local and global prototypes and prototypes are updated by an exponential moving average; at inference, each embedding is scored by the minimum fused cost to any prototype. The method is evaluated on MVTec AD and MVTec AD LOCO, reporting competitive or second-best average performance on LOCO and strong results on MVTec AD, with ablations on the spatial cost weight, number of prototypes, and batch size.","tokens_in":14038,"tokens_out":8056,"duration_ms":75463,"significance":"If the central claim is fully supported, the paper makes a useful contribution: it connects prototype-based anomaly detection to balanced optimal transport and introduces a simple structural cost that appears to help detect logical incoherencies. The code is released, the ablation study is informative, and the reported numbers are competitive with recent baselines. The main strength is that the central claim is falsifiable: the contribution of the structural constraint can, in principle, be separated into a learning-side and an inference-side effect. At present that separation is not performed, so the significance is conditional on the results of the additional experiment requested below.","major_comments":[{"comment":"The prototype update is written in Eq. (5) as p_i <- -η p_i + (1-η) Np/(B*H*W) Σ_k T*(k,i) z_k, but Algorithm 1 in Appendix B gives p_i <- η p_i + (1-η)(n*H_l*W_l)/(B*H_l*W_l) Σ_k T*(k,i) z_k. The sign of the first term differs. If Eq. (5) is taken literally, the update is not an exponential moving average and can drive prototypes away from the data; if the pseudocode is the intended update, Eq. (5) must be corrected. Because this is a central formula, the inconsistency must be resolved.","section":"3.1, Eq. (5) vs Appendix B Algorithm 1"},{"comment":"The ablation varies alpha jointly in training (Eqs. (1), (3), and (5)) and in inference (Eq. (6)), so the reported improvement for logical anomalies does not identify whether the gain comes from learning spatially organized prototypes or merely from applying the spatial penalty at test time. The abstract and Section 5 claim a benefit 'when learning the prototypes,' but this is not established by the current experiments. Please add a two-way experiment, for example training with alpha = 0 and scoring with alpha in {0, 0.3, 0.5} versus training with alpha = 0.3 and scoring with the same inference values; this directly tests whether local prototype learning contributes beyond the inference-side structural cost.","section":"4.3, Fig. 3b and Table 2"},{"comment":"There is a numerical inconsistency in the reported baselines: MVTec AD LOCO localization for PaDiM is 42.6 in Table 1 but the category-wise mean in Appendix A Table 3 is 52.1. The value 42.6 appears to be copied from the DRAEM row. Since Table 1 is used to support the 'second rank' claim, this error must be corrected and all affected averages recomputed.","section":"Table 1 vs Appendix A Table 3"},{"comment":"The values of the main hyperparameters (alpha = 0.3, n = 16, eta = 0.95, epsilon = 0.01, 50 epochs, batch size 64) are stated without a validation procedure. Since Fig. 3b shows that alpha strongly affects the trade-off between logical and structural anomaly detection, the reader cannot rule out that the headline numbers were obtained by selecting alpha on the test set. Please specify how each hyperparameter was chosen, for example via a fixed schedule, a held-out set of normal images, or a reported sensitivity range, and ideally report standard deviations over multiple runs.","section":"4.1, hyperparameters"}],"minor_comments":[{"comment":"The marginal constraint 'tT 1_{NHW} = nu' should read T^T 1_{NHW} = nu; as typeset, the transposition is confusing.","section":"3.1, Eq. (2)"},{"comment":"Eq. (1) defines the cost without the max-normalization that appears in Eq. (3); it would help to state explicitly that Eq. (1) is the conceptual cost and Eq. (3) is the normalized implementation.","section":"3.1, Eqs. (1) and (3)"},{"comment":"The uniform-marginal assumption is asserted without discussion. Given that normal industrial images often contain a dominant background, a sentence acknowledging this limitation, or a sensitivity check using reweighted marginals, would be useful.","section":"3.1, paragraph after Eq. (2)"},{"comment":"No error bars or multiple-seed results are reported. Given the small margins in Table 1 (for example PRADOT versus PaDiM/SPADE on LOCO localization), seed-averaged results with standard deviations would strengthen the ranking claims.","section":"4.1 and Table 1"},{"comment":"The pseudocode contains rendering artifacts, including garbled symbols in the input and initialization lines; these should be cleaned before publication.","section":"Appendix B"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read PRADOT. The first thing to know: it's a competent prototype-based UAD method with a genuinely new cost function, but the paper's main causal claim—that spatial regularization during prototype learning improves logical anomaly detection—is not supported by the reported ablation.\n\nWhat is new: the combination of optimal-transport prototype learning with a fused feature-and-spatial cost is not in the cited baselines like PatchCore or the few-shot OT methods. The authors show sensible ablations linking the spatial cost parameter α to logical versus structural anomalies, and they ship code. Performance is on par with strong baselines on MVTec AD and second on LOCO, which is a reasonable empirical result.\n\nSoft spots, in order of importance. First, the ablation in Section 4.3 varies α jointly in training and inference. The text in the abstract and conclusion says the structural constraint is enforced \"when learning the prototypes,\" but Fig. 3b and Fig. 4 show only the assignment cost at test time. The paper itself describes the effect as \"imposing a structural constraint on the embedding-to-prototype assignment\" (Section 4.3), which is an inference-side statement. To claim prototype learning benefits from α, you need at least one cell where training uses α=0 and inference uses α>0, or vice versa. As reported, the gain could come entirely from the test-time spatial penalty. That is a load-bearing gap, not a minor nit.\n\nSecond, Eq. (5) is inconsistent with the pseudocode in Appendix B. The main text shows pi ← -η pi + (1-η)... while the pseudocode and standard EMA use +η pi. One of them is wrong; this matters for reproducibility.\n\nThird, hyperparameters (α, n, η, ε) are set without any described validation procedure, and results are reported without error bars across seeds. This is common in the field but still a weakness for a paper making a quantitative claim about LOCO performance.\n\nThe uniform-marginal concern in Eq. (2) is real but minor. A dominant background could bias prototypes toward rare spatial locations, but this is a modeling assumption, not a fatal flaw.\n\nOverall, the method is plausible and the empirical work is honest, but the central attribution is unproven. The paper deserves a serious referee, mainly because the issue is fixable with a clean controlled ablation and the method itself is likely useful to the industrial UAD community. My recommendation: send to peer review, but ask the authors to separate training and inference α, fix Eq. (5), and report seeds.","headline":"PRADOT is a competent prototype-based UAD method with a genuinely new feature-spatial OT cost, but its advertised claim about spatial structure during prototype learning driving the gain is not established by the reported ablation.","tokens_in":14492,"tokens_out":1846,"would_cite":false,"duration_ms":20373,"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":"The paper proposes an unsupervised anomaly detection method, PRADOT, whose structural cost in optimal transport improves detection of logical image incoherencies.","keywords":["unsupervised anomaly detection","prototype learning","optimal transport","anomaly localization","logical anomalies","structural constraint","multi-scale features","entropic regularisation"],"falsifier":"Take a normal-image training set with a dominant background and rare foreground positions, train PRADOT twice, once with uniform marginals and once with marginals proportional to observed location frequency, and compare localisation of anomalies placed at rare positions: if the uniform-marginal model localises them markedly worse, the uniform assumption is the cause.","tokens_in":13624,"feed_emoji":"🔍","tokens_out":10236,"duration_ms":96788,"temperature":0.7,"pith_summary":"The paper proposes PRADOT, an unsupervised anomaly detection method that learns local and global prototypes from pre-trained image features using optimal transport. Its central claim is that adding a spatial cost to the feature-comparison cost during prototype learning and assignment captures the structural organisation of normal images, and that this improves detection of logical incoherencies such as misplaced objects. The method is evaluated on two industrial benchmark datasets, where it reports performance on par with compared baselines and ranks second on MVTec AD LOCO, with the structural cost shown to drive gains on logical anomalies. A sympathetic reader should care because it offers a single knob, the spatial-cost weight, that trades between detecting shape-level defects and detecting violations of expected layout, and it makes those failures visible through embedding-to-prototype assignment maps.","feed_headline":"Adding a spatial cost catches misplaced objects in images","feed_subtitle":"A feature-versus-position trade-off lets the detector flag logical incoherencies and ranks second on a public benchmark.","key_machinery":"The central object is the cost function $C(f,P) = (1-\\alpha)(1 - \\frac{\\langle z,p\\rangle}{\\|z\\|\\|p\\|}) + \\alpha \\|c-\\rho\\|^2$, which combines a cosine feature distance with an L2 spatial distance and is embedded in an entropic optimal transport problem between batch embeddings and prototypes. The named mechanism is the spatial or structural cost controlled by $\\alpha$: because prototypes occupy fixed grid coordinates and the transport problem couples features to those coordinates, increasing $\\alpha$ forces assignments to respect image layout. Prototype vectors are updated as weighted barycentres of assigned embeddings via an exponential moving average, after solving the entropically regularised optimal transport problem. The final anomaly map averages scores from global ($\\alpha = 0$) and local ($\\alpha > 0$) prototypes at two encoder scales.","core_discovery":"PRADOT builds a set of prototypes placed on the same spatial grid as encoder features, with several prototypes per grid cell. During training, an entropic optimal transport plan assigns each embedding to prototypes under uniform marginal distributions, and prototype vectors are updated by exponential moving average, so a prototype becomes a barycentre of the features assigned to it. The cost between an embedding and a prototype is $C(f,P) = (1-\\alpha)(1 - \\frac{\\langle z,p\\rangle}{\\|z\\|\\|p\\|}) + \\alpha \\|c-\\rho\\|^2$, where the second term penalises spatial displacement. With $\\alpha = 0$ the prototypes are purely feature-based or global; with $\\alpha > 0$ they become local, so an embedding whose feature matches a prototype far away pays a spatial penalty. At inference the minimum cost to any prototype is the anomaly score, and local and global maps at multiple scales are averaged. The claimed effect is that higher $\\alpha$ specifically improves detection and localisation of logical anomalies on MVTec AD LOCO, while a lower $\\alpha$ remains preferable for structural defects.","pith_inferences":["One testable extension is to replace the uniform embedding marginal with an empirical location-frequency marginal; if a dataset has dominant background regions, this should change which prototypes form and where localisation errors concentrate.","The $\\alpha$ trade-off suggests a per-category or adaptive $\\alpha$ selected on validation could keep the structural-defect performance on MVTec AD while preserving the logical-anomaly gains on LOCO.","The same spatial-assignment mechanism transfers naturally to medical imaging, where normal anatomy has strong layout regularity and the anomalies of interest are misplaced or duplicated structures; the uniform-marginal assumption would need revisiting there.","The assignment vectors from embedding to prototype can themselves be read as a saliency map for why a region is flagged."],"forward_implications":["Tuning $\\alpha$ lets one pipeline interpolate between detecting structural defects and detecting logical incoherencies, so a user can choose the operating point for the anomaly type that matters.","Because the assignment penalises spatial displacement, a patch that is normal in itself but appears in the wrong location receives a high anomaly score.","Averaging global and local prototype maps, and combining two feature scales, improves both image-level and pixel-level results over either alone.","The reconstructed-prototype visualisation provides a native explanation: abnormal regions are replaced by semantically coherent patches, which could also serve as an image-restoration side output.","The embedding-to-prototype assignment arrows give a per-image account of why a region was scored anomalous."],"supporting_citations":[{"why":"introduces the MVTec AD LOCO benchmark with logical anomalies and the AU-sPRO localisation metric that the method's core claim is evaluated on.","marker":"[1]"},{"why":"provides the MVTec AD dataset of structural defects and textures that the method is also benchmarked on.","marker":"[2]"},{"why":"supplies the entropic regularisation and fast solver that make the optimal transport problem tractable for large batches.","marker":"[8]"},{"why":"defines a structured optimal-transport barycentre framework that inspires the feature-plus-spatial cost and prototype update.","marker":"[35]"},{"why":"is the memory-bank prototype baseline whose MVTec results the paper compares against, particularly on the trade-off between logical and structural anomalies.","marker":"[25]"},{"why":"provides the uniform baseline results for all compared methods on both benchmarks.","marker":"[38]"},{"why":"motivates optimal-transport assignment to prototypes, equipartition, and the reliance on a sufficiently large batch-to-prototype ratio.","marker":"[5]"},{"why":"supplies the pre-trained ResNet architecture from which the multi-scale features are extracted.","marker":"[14]"}],"fun_headline_variants":["Spatial cost in prototypes catches misplaced image features","Prototypes with a spatial pull flag incoherent images","Optimal transport prototypes add a position check for anomalies","Balancing feature and spatial cost reveals logical defects"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every spatial location and every prototype should receive equal mass in the transport problem, which can misrepresent normal data with a dominant background or strongly non-uniform object placement.","fun_headline_variants_meta":{"raw":{"variants":["Spatial cost in prototypes catches misplaced image features","Prototypes with a spatial pull flag incoherent images","Optimal transport prototypes add a position check for anomalies","Balancing feature and spatial cost reveals logical defects"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001354,"raw_usage":{"total_tokens":5493,"prompt_tokens":935,"completion_tokens":4558,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":551,"completion_tokens_details":{"reasoning_tokens":4496}},"tokens_in":551,"tokens_out":4558,"duration_ms":34416,"temperature":1.0,"reasoning_tokens":4496,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:16:44.099241+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a normal-image training set with a dominant background and rare foreground positions, train PRADOT twice, once with uniform marginals and once with marginals proportional to observed location frequency, and compare localisation of anomalies placed at rare positions: if the uniform-marginal model localises them markedly worse, the uniform assumption is the cause.","supporting_citations":[{"cited_title":"In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)","cited_arxiv_id":null,"evidence_quote":"provides the MVTec AD dataset of structural defects and textures that the method is also benchmarked on."},{"cited_title":"In: Advances in Neural Information Processing Systems","cited_arxiv_id":null,"evidence_quote":"supplies the entropic regularisation and fast solver that make the optimal transport problem tractable for large batches."},{"cited_title":"In: Proceedings of the 36th International Conference on Machine Learning","cited_arxiv_id":null,"evidence_quote":"defines a structured optimal-transport barycentre framework that inspires the feature-plus-spatial cost and prototype update."},{"cited_title":"IEEE Transac- tions on Cybernetics54(5), 2720–2733 (2024)","cited_arxiv_id":null,"evidence_quote":"provides the uniform baseline results for all compared methods on both benchmarks."}],"review_version":1}