{"id":"bcc5b41b-a3be-4de2-a3c4-1e137c38095c","arxiv_id":"2412.04304","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":10,"one_line_summary":"3DzAL performs zero-shot 3D anomaly localization by combining contrastive patch learning, a normalcy classifier, and adversarial perturbation on pseudo-anomalies generated from task-irrelevant point clouds.","lead":"Researchers introduce a zero-shot setting for 3D anomaly detection, where the system trains only on normal objects from other categories. Their 3DzAL framework uses synthetic defects and patch-level classifiers to locate damage in unseen classes, beating two adapted baselines on MVTec 3D-AD.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA gains may be confounded by unequal auxiliary-data access: 3DzAL trains on pseudo-anomalies from 8 extra classes; BTF and 3DSR see only the single train class.","rationale":"The reader identified pseudo-anomaly transferability as the weakest assumption. I agree that assumption is important, but I see a more immediately load-bearing threat to the central claim in the evaluation protocol. The central claim is empirical: 3DzAL outperforms SOTA baselines with gains in every category. That claim cannot be fairly evaluated if the baselines are limited to one class while 3DzAL is allowed to use eight additional classes for pseudo-anomaly training. Section 4's leave-one-out setup gives 3DzAL access to the eight remaining classes per experiment; BTF and 3DSR receive only the single training class. Consequently, the reported gains in Tables 1 and 2 conflate algorithmic contribution with data budget. This is not an internal inconsistency but a fairness-of-comparison issue. The concrete test of building a BTF memory bank from all nine non-test classes would directly settle it. The paper's ablation in Table 3 further supports the concern: the contrastive loss, the component most tied to pseudo-anomaly transfer, contributes only about 0.1-0.5 AUPRO points, while the input-perturbation heuristic contributes the largest share of the gain. This does not change the reader's conditional verdict; it adds a specific condition: control for auxiliary-data access before accepting the SOTA claim. My assessment is therefore UNCHANGED with respect to the reader's CONDITIONAL verdict.","tokens_in":15400,"tokens_out":15723,"duration_ms":159259,"concrete_test":"Run the adapted BTF baseline with a memory bank built from the normal training patches of all 9 non-test classes (the same data 3DzAL uses), keeping all other BTF hyperparameters identical and using the same 90 train/test splits. Compute the mean pixel-level AUPRO over those splits. If the 9-class BTF equals or exceeds 3DzAL's mean, the reported SOTA gains are an artifact of unequal data access; if 3DzAL still wins, the concern is resolved. For completeness, also run 3DSR trained on all 9 non-test classes.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim that 3DzAL outperforms BTF and 3DSR with consistent AUPRO gains (Table 1) is not yet supported because the comparison does not control for data budget. In the leave-one-out protocol (Sec. 4), each of the 90 experiments gives 3DzAL the 8 non-train, non-test classes as task-irrelevant data for pseudo-anomaly synthesis, in addition to the one training class. The adapted baselines receive only that single training class. Thus 3DzAL sees nine classes of normal data; BTF and 3DSR see one. The reported gains could reflect this auxiliary-data advantage rather than the proposed cross-class pseudo-anomaly transfer. The paper's own ablation (Table 3) shows the contrastive loss Lcon adds at most about 0.2 AUPRO, suggesting the mechanism is not the main driver, yet the data-budget asymmetry affects every reported result. A fair zero-shot comparison must give baselines access to the same nine non-test classes, e.g., a BTF memory bank containing all of them.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the task of zero-shot 3D anomaly detection and localization, where the training class and test class do not overlap, and proposes 3DzAL. 3DzAL uses a randomly initialized, untrained CNN on ordered xyz point clouds to generate pseudo-anomaly patches from eight task-irrelevant classes, trains a PointNet++ feature extractor with contrastive and disentanglement losses, trains a normalcy classifier, and combines a PatchCore-style distance score with an adversarially perturbed classification score. Experiments on MVTec 3D-AD follow a leave-one-out protocol over 10 classes (90 train/test pairs) and report pixel-level AUPRO and image-level AUROC against BTF and 3DSR adapted to the zero-shot setting, with ablations over components, pseudo-anomaly types, and CNN initialization.","tokens_in":15733,"tokens_out":8324,"duration_ms":81956,"significance":"The proposed task is practically motivated and the paper is one of the first to address it. If the reported comparisons were controlled, the framework would be a solid zero-shot baseline. The observation that a random CNN on xyz input produces activation maps that highlight anomalous-looking regions (Fig. 3) is interesting and potentially reusable beyond anomaly detection. Strengths include the exhaustive 90-experiment leave-one-out evaluation, explicit ablations (Tables 3-6), and comparison of model/memory-bank sizes. However, the main empirical claim is currently confounded by unequal access to auxiliary data and by unreported final-score weights, so the significance cannot be fully assessed until those issues are addressed.","major_comments":[{"comment":"The leave-one-out protocol gives 3DzAL access to eight task-irrelevant classes for pseudo-anomaly generation in addition to the single training class, while the adapted BTF and 3DSR baselines are trained or built from that single training class only. The reported AUPRO/AUROC gains are therefore not attributable to the proposed cross-class transfer mechanism; they may instead reflect a 9-class versus 1-class data budget. Please rerun BTF and 3DSR with the same nine non-test classes (e.g., a BTF memory bank built from all nine classes and 3DSR trained on all nine) and compare, or ablate 3DzAL with reduced auxiliary data, so that the comparison controls for data access.","section":"Section 4, Experimental setting; Tables 1 and 2"},{"comment":"The final anomaly score uses weights wd and wc, but their values are never reported and no sensitivity analysis is provided. Since the reported AUPRO is computed from this weighted sum, the unspecified weights make the main tables non-reproducible and leave open the possibility of per-category tuning. Please report the values used for wd and wc and demonstrate that the conclusions are stable over a range of these weights.","section":"Section 3.4, Eqs. (11)-(12); Implementation details"},{"comment":"The 'inductive bias' of the random CNN is a core component of the method, but the paper does not compare pseudo-anomaly generation against a trivial baseline of randomly selected patches. Table 5 only compares random versus ImageNet-initialized CNN weights; without a random-patch control, the claim that the random CNN localizes anomaly-relevant regions (contribution 3) is not quantitatively supported. Please add such a control and, if possible, quantify the agreement between the activation mask A and ground-truth anomaly locations on task-irrelevant data.","section":"Section 3.2 and Table 5"},{"comment":"The paper explicitly states that it 'intentionally do[es] not make use of any existing zero-shot AD work' and therefore omits 2D zero-shot anomaly-detection baselines. Because 3DzAL has an RGB branch and the dataset provides RGB images, comparisons with simple zero-shot 2D methods (e.g., ImageNet-feature distance or CLIP-based scores on RGB) are feasible and necessary to support the statement that the method outperforms 'state-of-the-art anomaly detection and localization' in a zero-shot setting. Please add at least one such baseline or justify why RGB-only zero-shot methods are outside the comparison scope.","section":"Section 2 and Section 4, Baselines"},{"comment":"No error bars or statistical significance are reported for the 90 individual experiments, and several key hyperparameters (tau, negative-to-positive patch ratio, T, eta, b, patch size, epochs) are fixed without sensitivity analysis. Given the modest per-component gains in Table 3 (e.g., the contrastive loss adds no more than about 0.2 AUPRO in the reported settings), reporting variance over at least three random seeds and a sensitivity sweep for the most critical hyperparameters (tau, eta, b, and wd/wc) would materially strengthen the claim that the improvements are robust.","section":"Tables 1-3 and Implementation details"}],"minor_comments":[{"comment":"There are typographical artifacts such as 'W e' and 'Y un Fu' in the author block; these should be corrected.","section":"Abstract and author list"},{"comment":"The phrase 'reciprocal second, third, and fourth layer output' should be 'respective second, third, and fourth layer outputs', and 'comparitively' is a typo for 'comparatively'.","section":"Section 3.2"},{"comment":"The term 'KNN Gaussian Blurring' is unclear; if this refers to a kernel/Gaussian blur used to smooth the score map, it should be stated explicitly and distinguished from the nearest-neighbor parameter b.","section":"Section 3.4"},{"comment":"Reference [22] is listed as '3D-ADS. 2022.' with no title or venue; it should be completed.","section":"References"},{"comment":"The text says the ablations are run on bagel, dowel, and foam, but Table 4 reports bagel, potato, and rope; please clarify which training classes were used for each ablation table.","section":"Section 4 and Table 4"},{"comment":"The row/column structure reports means over the nine test classes per training class, but the captions say 'in all categories'/'in most categories'; please specify that these statements refer to the per-training-class means, since the individual test-class entries are not compared separately in the table.","section":"Section 4, Tables 1 and 2"},{"comment":"The definition of bp(xtest|w) as max{p(xtest|w), 1-p(xtest|w)} makes the sign of the perturbation step hard to follow; please clarify the predicted class and state explicitly why adding -eta times the gradient lowers the confidence of the predicted class.","section":"Section 3.4, Eq. (8)"},{"comment":"Please state how the 'ordered' 3D point cloud is obtained from the MVTec 3D-AD data (e.g., whether it is the native organized point cloud or the result of a projection/unprojection step), as this is needed for reproducibility.","section":"Implementation details"}],"recommendation":"major_revision","confidential_remarks":"The data-budget asymmetry between 3DzAL and the baselines is the most serious concern; if the authors rerun the baselines with the same auxiliary-class access and the reported gains persist, the paper would be suitable for publication after additional reproducibility details. I would also ask the editor to ensure the novelty claim about random-CNN inductive bias is backed by a random-patch baseline, and to treat the 'intentional' omission of zero-shot 2D baselines in Section 2 as a scope choice rather than a substitute for comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one for the problem definition and the random-CNN empirical observation, not for the SOTA claim as stated. The new setting—training on one class, testing on a different class with no target-class normal data—is clearly defined and practical, and the paper is the first to point at it for 3D. The finding that a randomly initialized ResNet's activation maps on ordered xyz tensors highlight defect-like regions (better than ImageNet-pretrained, per Table 5) is genuinely interesting and should survive as a standalone observation.\n\nWhat the paper does well: 90 leave-one-out experiments, ablations on every component, a 2-class generalization check, and a limitation statement about low-performing pairs. The pipeline is assembled from known pieces (PatchCore memory bank, FPFH, PointNet++, contrastive learning) but the combination is reasonable and the ablations show each piece contributes something.\n\nThe soft spot is the comparison in Tables 1 and 2. The protocol gives 3DzAL access to nine classes of normal data—the single train class plus eight task-irrelevant classes used to synthesize pseudo-anomalies—while BTF and 3DSR only see the single train class. So the reported AUPRO gains (2–7 points) could be a data-budget effect, not evidence that cross-class pseudo-anomaly transfer works. The stress-test note is right, and I checked the setup in Sec. 4: there is no control. To make the claim stick, the authors need to give baselines the same auxiliary data, e.g. a BTF memory bank over all nine classes (or at least an ablation with one auxiliary class for 3DzAL). The ablation in Table 3 also shows Lcon alone adds at most 0.2 AUPRO, which makes me think the pseudo-anomaly contrastive mechanism is not the main driver; the advantage may be mostly from extra data and the classifier.\n\nMinor issues: wd/wc are never specified; no error bars anywhere (some category numbers, e.g. 3DSR on foam/cookie, are suspiciously low); no code release. The deliberate exclusion of 2D zero-shot AD baselines is understandable but should be accompanied by at least a discussion or an adapted 2D baseline, since 3DzAL also uses an RGB branch from ImageNet which somewhat contradicts the \"no pretrained model\" claim in contribution 2.\n\nWho benefits: researchers working on 3D anomaly detection and anyone interested in inductive biases of random networks. It deserves a serious referee; the task and the random-CNN observation are valuable even if the empirical SOTA claim needs rework. I'd accept for review with a request to fix the data-budget confound, report weights/error bars, and release code.","headline":"Defines a useful zero-shot 3D AD task and shows a plausible pipeline, but the headline SOTA comparison is confounded by asymmetric data access.","tokens_in":16232,"tokens_out":2132,"would_cite":true,"duration_ms":20761,"reading_group":"yes","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 claims 3D anomaly localization can transfer to unseen classes without target-class normal data, using pseudo-anomalies from a random CNN's attention on task-irrelevant point clouds.","keywords":["zero-shot anomaly detection","3D anomaly localization","point cloud","pseudo anomaly generation","contrastive learning","random CNN inductive bias","normalcy classifier","adversarial perturbation"],"falsifier":"Run the same pipeline but replace the random-CNN point selection with the same number of uniformly random points; if pixel-level AUPRO does not drop noticeably, the claimed inductive bias is not what drives transfer.","tokens_in":15198,"feed_emoji":"🔍","tokens_out":6817,"duration_ms":63488,"temperature":0.7,"pith_summary":"This paper identifies a new problem in industrial inspection: localizing anomalies in 3D point clouds of an object class for which no normal training data exist, because the training and testing classes do not overlap. It proposes 3DzAL, a patch-level contrastive learning framework that trains on normal data from one class and on pseudo anomalies synthesized from task-irrelevant classes, then tests on a disjoint class. The central move is using a randomly initialized, untrained CNN to highlight locations of interest in xyz point clouds; those points are attached to or removed from normal patches to create fake defects. Combined with a normalcy classifier and an adversarial perturbation of the classifier input, the framework claims to outperform the previous state of the art in both pixel-level and image-level zero-shot anomaly detection on the standard ten-class benchmark. If the claim holds, a factory could deploy off-the-shelf defect localization for new 3D parts without collecting or sharing normal samples of those parts.","feed_headline":"3DzAL finds defects in object classes it never trained on","feed_subtitle":"Random-CNN pseudo anomalies plus contrastive learning make anomaly localization transfer to unseen 3D classes.","key_machinery":"The load-bearing object is the pseudo-anomaly generation module. Given ordered xyz data from task-irrelevant classes, an untrained, randomly initialized CNN produces multi-scale activation maps; the highest-activation $\\tau$ percent of points are treated as places of interest. Those points are attached to normal patches (adding-point anomalies) or a random patch has points removed (removing-point anomalies), in a 1:1 ratio. These pseudo anomalies feed a patch-level contrastive loss that pulls normal patch features together and pushes pseudo-anomaly features away, plus a representation-disentanglement loss that makes the learned 3D features complementary to FPFH features. A PointNet++ normalcy classifier is trained on the same normal-versus-pseudo-anomaly pairs with binary cross-entropy, and at test time its input is shifted against the gradient of its own softmax score before the classification probability is used as an anomaly score. The final score is a weighted sum of the memory-bank distance score and this perturbed classification score.","core_discovery":"3DzAL establishes that the normal-versus-abnormal distinction can be learned class-agnostically. On the ten-class 3D industrial inspection benchmark, the paper systematically leaves one class out: it trains on the normal data of one class, uses the remaining eight as pseudo-anomaly sources, and tests on a ninth unseen class, covering 90 train/test pairs. The learned representation combines RGB features, handcrafted FPFH features, and PointNet++ features trained by contrastive learning against pseudo anomalies; a separate normalcy classifier scores patch abnormality, and the final score fuses a memory-bank distance with the classifier output on adversarially perturbed patch coordinates. The paper reports that 3DzAL beats the two strongest prior methods (BTF and 3DSR) on every category for pixel-level AUPRO and on most categories for image-level AUROC, with the largest per-category gains on localization. It also reports that the random, untrained CNN's attention on xyz data localizes points of interest better than an image-pretrained CNN for this task.","pith_inferences":["Because the pseudo anomalies are class-agnostic geometric edits (adding and removing point clusters), the same generation recipe could be recycled for other 3D inspection tasks, such as detecting deformation in non-rigid objects, where the notion of 'normal shape' is class-specific.","A directly testable extension is to increase the diversity of task-irrelevant classes or add synthetic CAD data; if the transfer gains grow with pseudo-anomaly diversity, the core mechanism is confirmed as cross-class geometric contrast.","The adversarial perturbation on the classifier input is a test-time sharpening step; one could check whether it also improves the classical same-class setting, which would separate its value from zero-shot transfer specifically.","The fixed score weights $w_d$ and $w_c$ are set once in the paper; a probing experiment could check whether per-target-class tuning of these weights changes the ranking across categories, since the current averages may mask category-dependent sensitivity."],"forward_implications":["Anomaly localization can be done off-the-shelf for a new object class without collecting or sharing its normal data, which removes a practical barrier when target-class data are private or export-controlled.","The randomly initialized CNN's attention on xyz data is a free prior for 3D geometric interest points; it can be used anywhere a cheap saliency map over point clouds is needed.","The method's training signal depends only on generic geometric edits to normal patches, so the same recipe should transfer to other 3D anomaly benchmarks beyond the ten-class set studied here.","Because the learned features are complementary to FPFH, the framework does not need a 3D-pretrained model; the only pretrained component is the standard RGB backbone.","Fusing the distance score with the perturbed classifier score improves both localization and detection, suggesting that the two scores capture partially independent evidence."],"supporting_citations":[{"why":"Establishes that a randomly initialized CNN has an inductive bias to focus on objects in 2D images, the basis for extending this to 3D xyz data.","marker":"[7]"},{"why":"First used untrained-CNN inductive bias to synthesize 2D pseudo anomalies from task-irrelevant data, the direct predecessor of the 3D pseudo-anomaly module.","marker":"[1]"},{"why":"Supplies the memory-bank feature storage, coreset selection, and distance-based anomaly scoring that 3DzAL builds on for both training and testing.","marker":"[34]"},{"why":"BTF is the 3D adaptation of the memory-bank paradigm and the main zero-shot baseline that 3DzAL must outperform.","marker":"[23]"},{"why":"3DSR is the current state-of-the-art classical 3D anomaly detection method; adapting it to zero-shot is the strongest comparison and the failure case motivating the new approach.","marker":"[53]"},{"why":"Provides the handcrafted FPFH features used in the FPFH branch and as the target of the representation-disentanglement loss.","marker":"[36]"},{"why":"Contributes the MVTec 3D-AD dataset and the AUPRO/AUROC evaluation protocol used for all experiments.","marker":"[3]"},{"why":"PointNet++ is the backbone for both the 3D feature extractor and the normalcy classifier.","marker":"[32]"}],"fun_headline_variants":["Zero-shot 3D anomaly detection: find defects in unseen classes","3DzAL detects anomalies without training on target class","No target data? 3DzAL still localizes 3D defects","Contrastive pseudo-anomalies enable zero-shot 3D defect finding","3D anomaly localization that transfers to new object types"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole training signal rests on the assumption that pseudo-anomaly patches, built from points flagged by a randomly initialized CNN on task-irrelevant xyz data plus added or removed points, resemble the anomalies of the never-seen target class closely enough for the learned normal-versus-abnormal distinction to transfer.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot 3D anomaly detection: find defects in unseen classes","3DzAL detects anomalies without training on target class","No target data? 3DzAL still localizes 3D defects","Contrastive pseudo-anomalies enable zero-shot 3D defect finding","3D anomaly localization that transfers to new object types"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000176,"raw_usage":{"total_tokens":1302,"prompt_tokens":974,"completion_tokens":328,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":237}},"tokens_in":590,"tokens_out":328,"duration_ms":3452,"temperature":1.0,"reasoning_tokens":237,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:32:45.685221+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same pipeline but replace the random-CNN point selection with the same number of uniformly random points; if pixel-level AUPRO does not drop noticeably, the claimed inductive bias is not what drives transfer.","supporting_citations":[{"cited_title":"A random CNN sees objects: One inductive bias of cnn and its applications","cited_arxiv_id":null,"evidence_quote":"Establishes that a randomly initialized CNN has an inductive bias to focus on objects in 2D images, the basis for extending this to 3D xyz data."},{"cited_title":"Cross-domain video anomaly detection without target domain adaptation","cited_arxiv_id":null,"evidence_quote":"First used untrained-CNN inductive bias to synthesize 2D pseudo anomalies from task-irrelevant data, the direct predecessor of the 3D pseudo-anomaly module."},{"cited_title":"Towards total recall in industrial anomaly detection","cited_arxiv_id":null,"evidence_quote":"Supplies the memory-bank feature storage, coreset selection, and distance-based anomaly scoring that 3DzAL builds on for both training and testing."},{"cited_title":"Back to the feature: Classical 3d features are (almost) all you need for 3d anomaly detection","cited_arxiv_id":null,"evidence_quote":"BTF is the 3D adaptation of the memory-bank paradigm and the main zero-shot baseline that 3DzAL must outperform."},{"cited_title":"Cheating depth: Enhancing 3d surface anomaly detection via depth simulation","cited_arxiv_id":null,"evidence_quote":"3DSR is the current state-of-the-art classical 3D anomaly detection method; adapting it to zero-shot is the strongest comparison and the failure case motivating the new approach."},{"cited_title":"Fast point feature histograms (fpfh) for 3d registration","cited_arxiv_id":null,"evidence_quote":"Provides the handcrafted FPFH features used in the FPFH branch and as the target of the representation-disentanglement loss."},{"cited_title":"The MVTec 3D-AD dataset for unsupervised 3D anomaly detection and localization","cited_arxiv_id":null,"evidence_quote":"Contributes the MVTec 3D-AD dataset and the AUPRO/AUROC evaluation protocol used for all experiments."},{"cited_title":"Pointnet++: Deep hierarchical feature learning on point sets in a metric space","cited_arxiv_id":null,"evidence_quote":"PointNet++ is the backbone for both the 3D feature extractor and the normalcy classifier."}],"review_version":1}