Pith. sign in

REVIEW 4 major objections 5 minor 55 references

The paper's central claim is that object-centric state tracking, not language reasoning, is the bottleneck in industrial video anomaly detection with vision-language models.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

An object-centric, training-free agentic pipeline that tracks object state changes and reasons over them with a vision-language model achieves strong video-level AUROC on Phys-AD, LiquidAD, and IPAD, while producing interpretable anomaly reports.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Object-centric state tracking is a real improvement for IVAD, but the paper overclaims SOTA and the IPAD protocol contradicts its no-domain-knowledge promise. the 4 major comments →

arxiv 2607.18142 v1 pith:S3VOQJXI submitted 2026-07-20 cs.CV cs.AIcs.CLcs.MA

O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning

Mei Yuan , Qi Long , Qifeng Wu , Zhenyang Li , Yizhou Zhao , Lei Wang , Yang Liu , Min Xu This is my paper
classification cs.CV cs.AIcs.CLcs.MA
keywords industrial video anomaly detectionvision-language modelsobject-centric state trackingtraining-free agentic frameworkchain-of-thought reasoningopen-ended anomaly reportingspatiotemporal tubeletsvisual verification
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Industrial video anomaly detection demands catching faults that appear as gradual object transformations—a toothpaste tube that deforms before leaking, a screw that backs out before its head strips. The paper's central claim is that current vision-language models fail at this not because they reason poorly, but because they are given the wrong evidence: whole frames or captions instead of per-object state histories. To test this, the authors build a training-free, three-stage pipeline that discovers objects, tracks each object's state changes over time, and then runs a structured chain-of-thought over those state trajectories. On three industrial benchmarks the pipeline reports higher video- and frame-level anomaly-detection scores than fine-tuned traditional detectors and frontier vision-language models, while also emitting open-ended reports with anomaly type, affected object, frames, severity, and cause. If true, the implication is that grounded object-level evidence can substitute for training data and domain-specific knowledge in industrial inspection.

Core claim

The paper claims that the bottleneck in VLM-based industrial video anomaly detection lies in object-level evidence construction, not in language reasoning. It proposes O-VAD, a training-free framework that 'grounds, tracks, and reasons': it discovers and segments objects with a vision-language model, builds per-object temporal state trajectories through spatiotemporal tubelet tracking with track recovery after transformations, and then feeds the accumulated state-change events into a six-step chain-of-thought that separates expected process outcomes from failures and verifies borderline candidates visually. The paper reports state-of-the-art performance on three IVAD datasets at both video a

What carries the argument

The central evidence object is the per-object state trajectory: for each tracked object k, a sequence of masks and natural-language state descriptions across frames, along with open-ended state-change event tuples carrying start/end frames, change type, cause, description, severity, and affected object index. Stage 2 constructs these trajectories by partitioning the video into tubelets, propagating masks with a video segmentation model, and recovering post-transformation tracks via spatial-proximity and semantic-consistency gates before querying a VLM for free-form state changes. Stage 3 reasons over the trajectories with a cascaded chain-of-thought and applies confidence-tiered visual verif

Load-bearing premise

The load-bearing assumption is that Stage 2's VLM-generated state-change events are complete and faithful evidence of everything relevant in the video; if a state change is missed or hallucinated, the later reasoning and verification—which use the same VLM family—will confidently report a wrong conclusion, and the paper's own failure cases (clip cannot press, stuck button, restricted servo, degaussed magnet) show this happening.

What would settle it

Run the pipeline's Stage 3 on the Phys-AD failure cases while injecting oracle state-change events that include the missing action ('pressing tool contacted the clip but the clip did not move', 'button did not depress'). If the VLM still returns 'no anomaly', the bottleneck is reasoning, not evidence construction; if it now detects the faults, O-VAD's misses are caused by incomplete state detection, confirming the paper's claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Object state tracking, not better prompting or bigger reasoning, is what lifts VLM performance: removing it drops recall to near zero on three of four ablation categories, so any VLM-based industrial detector needs an explicit per-object temporal evidence layer.
  • Training-free detection can match or exceed fine-tuned traditional VADs on these benchmarks, suggesting per-class 'one-class-one-model' training is not a prerequisite for reliable industrial anomaly detection.
  • The open-ended output—anomaly type, affected object, frame range, severity, and causal analysis—makes the report directly actionable for root-cause diagnosis, addressing the interpretability requirement of industrial inspection.
  • Anomalies defined by invisible properties or missing actions (restricted servo angle, stuck button, degaussed magnet, clip unable to press) remain a boundary; the paper identifies few-shot in-context specification examples and lightweight domain priors as the next step.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A failure mode the paper reports—no detected state change leads to 'no anomaly' even when the ground truth is a missing expected action—implies that expectation modeling is the next logical extension: if the VLM were asked to predict what should happen before comparing with observed state changes, absent actions could become positive evidence without any training.
  • The per-object state-trajectory representation is a reusable intermediate beyond the paper's VLM reasoning: it could feed lightweight classifiers, or its event density (which clusters at cycle boundaries on periodic normal processes) could serve as a label-free anomaly heuristic for monitoring systems.
  • The reported LLM-judge split—a fluent baseline scores higher on explanation metrics while O-VAD wins detection—suggests evaluation of explainable anomaly detection should condition explanation quality on correctness; otherwise fluent but wrong narratives look better than correct but less polished ones.
  • Stratifying benchmarks into 'perception-solvable' and 'spec-required' anomalies would make cross-method comparison fairer, since the paper's failure analysis indicates no visual-evidence method can resolve the latter without reference specifications.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper presents O-VAD, a training-free agentic framework for industrial video anomaly detection that operates in three stages: (1) VLM-grounded object discovery with SAM3, (2) object-centric state tracking via spatiotemporal tubelets (CropFormer + SAM2) plus open-ended VLM state-change detection, and (3) a six-step chain-of-thought anomaly reasoning stage with confidence-based visual verification. The framework is evaluated on Phys-AD, LiquidAD, and IPAD, reporting video-level, frame-level, and type-level metrics against traditional trained VADs, direct-prompting VLMs, and two agentic baselines. The central claim is that object-level state tracking, rather than language reasoning, is the bottleneck in industrial VLM-based video anomaly detection, and that O-VAD achieves state-of-the-art performance without training, predefined taxonomies, or domain-specific context injection.

Significance. If the claims hold, O-VAD represents a useful step toward explainable, open-set industrial video anomaly detection. The paper's strengths include a fully training-free pipeline, explicit moves from frame-level captioning to object-centric state trajectories, interpretable anomaly reports with grounded frame ranges and causal analyses, and a relatively extensive experimental appendix including human evaluation and failure-case analysis. The method is timely and the qualitative examples (e.g., leakage and screw-stripping cases) are compelling. However, several load-bearing claims are not supported as stated: the IPAD evaluation injects a normal reference clip into O-VAD's prompt, the reported BERTScore claim contradicts Table 1, the intermediate state-change evidence is produced and consumed by the same VLM family without independent validation, and the abstract's 'state-of-the-art at both video- and frame-levels' is not accurate for LiquidAD frame-level AUROC nor applicable to Phys-AD, which has no frame-level metric. These issues are fixable but require substantive revision.

major comments (4)
  1. The IPAD evaluation violates the paper's core 'no domain knowledge / no context injection' claim. Appendix B.2 states that for IPAD, O-VAD injects 'one normal reference case (three evenly sampled frames) as additional model input' and replaces the 'expectation' part of the prompt with reference to those frames. This is dataset-specific, test-time context injection. The IPAD baselines in Table 1 (Qwen3-VL-32B, GPT-5, URF-ZS-HVAA) do not receive this reference, making the IPAD comparison unfair. Since the abstract explicitly says O-VAD works 'without injecting domain knowledge as context for test-time inference', the IPAD results cannot be used to support the headline claim. Please rerun IPAD without reference injection for O-VAD, and/or give the same reference frames to all baselines, and transparently report both settings. At minimum, the abstract and contributions must qualify the claim
  2. The text states: 'It also yields the most semantically faithful descriptions, with the best type-level BERTScore (0.803) among training-free methods.' Table 1 lists GPT-5 with type-level BERTScore 0.878, which is higher than O-VAD's 0.803. GPT-5 is marked as training-free ('*'). Thus the claim 'best BERTScore among training-free methods' is directly contradicted by the paper's own table. Either the comparison is restricted to a subset that is not stated, or the sentence is wrong. Please correct the claim or clarify the metric construction. This matters because the paper uses BERTScore as evidence for the quality of O-VAD's open-ended reports.
  3. The evidence chain is not fully external. Stage 2's state-change event tuples (Eq. 8) are generated by GPT-5 from masked frame pairs; Stage 3's reasoning and the post-verification confidence (Eqs. 9-10) are produced by the same model family. The report can therefore be internally self-consistent without independent pixel-level grounding. The label-free axes in C.1 (count, content, pattern) and the human evaluation in C.5 assess final reports, not the intermediate state-change events. The failure cases in §C.4 (clip unable-to-press, stuck button, restricted servo, degaussed magnet) show that when no state change is detected, O-VAD reports 'no anomaly' — exactly the failure mode that would result from hallucinated or incomplete state events. Please validate Stage 2 directly: e.g., have human annotators or an independent VLM verify a sample of state-change events against the video frames, a
  4. The abstract claims 'state-of-the-art performance on three IVAD datasets at both video- and frame-levels.' This is not supported. Phys-AD has no frame-level metric in the paper; Table 4 lists its annotations as Video and Object only. On LiquidAD frame-level AUROC, O-VAD scores 0.512, lower than the trained S3R baseline (0.625) and comparable to URF-ZS-HVAA (0.506). The claim is accurate only for video-level AUROC averages (and for some frame-level recall/F1 values on LiquidAD/IPAD). Please qualify the claim to the specific metrics and levels that are actually demonstrated, and describe the Phys-AD per-category results in Table 2 as 'best on average' rather than universal SOTA, since MNAD.p or S3R win many individual categories.
minor comments (5)
  1. The O-VAD report in the bottom row appears to contain state-change entries from a different video (references to 'Metallic tube (likely a toothpaste or ointment tube)' and repeated deformation of a lower edge) in a clip-pressing example. This looks like a copy-paste artifact from another case. Please replace with the correct tracking output.
  2. Typo: 'withy open-ended state change detection' should be 'with open-ended state change detection'.
  3. In the GPT-5 output for the hinge example, there is an obvious typo: 'nThehinge should remain stationary if fixturedohinge deformation' — appears to be a text-merge error. Please correct.
  4. The caption says best/2nd-best are highlighted in / but the table rendering may not show these colors in all PDF viewers. Consider adding bold or a legend that survives monochrome printing.
  5. The Stage-3 verification stride is dataset-specific (60/30/25 for Phys-AD/LiquidAD/IPAD). This is a legitimate implementation detail, but it is a per-dataset choice that should be disclosed in the main text where 'parameter-free' or 'no dataset adaptation' is implied.

Circularity Check

2 steps flagged

IPAD evaluation injects the reference normal that defines the target, and the final verification loop is performed by the same VLM that generated the evidence.

specific steps
  1. self definitional [Appendix B.2 (IPAD protocol); Section A.1 (dataset description)]
    "Specifically, when prompting the VLM for anomaly reasoning, we inject one normal reference case (three evenly sampled frames) as additional model input and replace the "expectation" part of prompt with pointing to the reference frames accordingly."

    The paper claims the method is "free of domain-specific knowledge" and works "without injecting domain knowledge as context for test-time inference" (Abstract). On IPAD, however, the test-time prompt includes a dataset-specific normal reference clip. IPAD's ground truth is defined as "deviations in object appearance ... and spatial location relative to reference normal samples" (Section A.1). Thus the input supplies the very reference that defines the target: the model is asked to detect deviations from an injected normal example. This makes the IPAD evaluation a comparison-to-given-reference task rather than a no-context test, and the baselines in Table 1 were not given the same reference frames, making the IPAD SOTA claim unsupported as stated.

  2. other [Section 3.3 (Eq. 8) and Section 3.4 (Eqs. 9-10)]
    "Each detected state change event e is represented as a tuple: e=(tstart, tend, type, cause, desc, sev, k) ... (verified, cver) = φVLM({It}t∈evidence, c, a)."

    The only anomaly evidence (state-change events, Eq. 8) is generated by φVLM (GPT-5) in Stage 2, and Stage 3's chain-of-thought reasoning and post-hoc verification (Eq. 9) are performed by the same model family. The final confidence in Eq. 10 is a product of the VLM's initial confidence and its own verification confidence. Consequently, a hallucinated or incomplete state-change list can be 'verified' by the same model without any independent pixel-level evidence, and the report can be internally self-consistent while wrong. This is a closed evidentiary loop rather than an externally grounded chain, weakening the claim that reports are 'grounded in concrete, per-object evidence'.

full rationale

The paper's central contribution is an object-centric state-tracking pipeline that is claimed to need no domain-specific knowledge or test-time context. The IPAD experiments contradict that claim directly: the method receives a normal reference clip as part of the prompt, and IPAD anomalies are defined as deviations from exactly that kind of reference. The IPAD column therefore cannot be used to support the 'no context' headline, and the baseline comparison is unfair because direct-prompting baselines do not receive the same reference. This is a clear, quotable flaw, though it affects one of three benchmarks and does not by itself invalidate the Phys-AD or LiquidAD results. A second concern is that the entire evidence chain—state-change detection, anomaly reasoning, and post-verification—is executed by the same VLM (GPT-5). The final confidence is a product of the VLM's own initial and verification confidence values, so the pipeline can be self-consistent without external grounding. This is not a logical derivation from inputs to outputs, but it is a closed loop that reduces the strength of the 'grounded' claim. Weighed together, there is partial circularity in the IPAD adaptation and a self-referential verification loop, but the core Phys-AD/LiquidAD results still have independent benchmark content, so a moderate score (4) is appropriate.

Axiom & Free-Parameter Ledger

9 free parameters · 5 axioms · 0 invented entities

The central claim rests on a stack of inherited and hand-chosen thresholds, on the assumption that VLM-generated state changes are faithful, and on a dataset-specific IPAD prompt modification. No new physical entities are introduced. The free parameters are not fitted constants in a physics derivation, but they materially shape the reported AUC and are not all rigorously searched or released as code.

free parameters (9)
  • VLM detection confidence threshold tau_conf,1 = 0.1
    Stage 1 permissive recall threshold; chosen by authors, affects object discovery and all downstream tracking.
  • Tubelet coverage threshold tau_coverage = 0.25
    Inherited from TubeletGraph [35]; governs when new tubelets are initiated.
  • Spatial proximity gate tau_prox = 0.3
    Track recovery gate for post-transformation objects; inherited from TubeletGraph grid search.
  • Semantic consistency gate tau_sem = 0.7
    CLIP similarity threshold for track recovery; validated as near-optimal in TubeletGraph and adopted here.
  • High-confidence bypass tau_hi = 0.8
    Stage 3 confidence gating; chosen by authors and affects which anomalies skip verification.
  • Low-confidence discard tau_lo = 0.2
    Stage 3 confidence gating; chosen by authors and affects which candidates are discarded.
  • Post-verification retention tau_conf = 0.3
    Stage 3 confidence gating; chosen by authors and affects final anomaly set.
  • Stage-2 content-adaptive FPS = [2,10], default 3
    Frame sampling for state-change detection; changes temporal granularity of evidence.
  • Stage-3 dataset-specific stride = 60/30/25 for Phys-AD/LiquidAD/IPAD
    Sampling stride for visual verification frames; affects which evidence frames are seen.
axioms (5)
  • domain assumption VLM internalized commonsense suffices to detect and describe object state changes and to distinguish expected process behavior from anomalies
    The whole method is training-free; if GPT-5 cannot reliably perceive small state changes, Stage-2 evidence and Stage-3 verdicts inherit the failure. Failure cases in §C.4 confirm this boundary.
  • domain assumption SAM3, CropFormer, and SAM2 provide complete and identity-stable object segments/tubelets through major transformations
    Stages 1 and 2 depend on these; the paper states that a missed object or track is irrecoverable (§A.2).
  • domain assumption Object state-change events are more reliable than frame-level captions for anomaly reasoning
    This is the core premise of the paper (§1, §3); if false, the design collapses to caption-level baselines.
  • ad hoc to paper For IPAD, injecting one normal reference case into the Stage-3 prompt transfers the dataset's reference-based anomaly definition without domain-specific knowledge
    §B.2 modifies the CoT prompt with reference frames; this is a dataset-specific adaptation that the abstract's 'no domain-specific knowledge' claim does not disclose.
  • domain assumption Anomaly labels and benchmark metrics (AUROC, BERTScore, LLM-as-judge, human questionnaire) are valid for open-ended outputs
    Evaluation relies on BERTScore and an LLM-as-judge protocol; the paper itself acknowledges LLM-judge fluency bias in §C.6.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning." pith.science (2026). https://pith.science/paper/S3VOQJXI

@misc{pith2026260718142,
  author       = {Pith},
  title        = {Pith review of: O-VAD: Industrial Video Anomaly Detection through Object-Centric Tracking and Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S3VOQJXI}},
  note         = {Machine review of arXiv:2607.18142}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Industrial Video Anomaly Detection (IVAD) aims to identify anomalous objects and events in an industrial process, which is crucial for modern manufacturing and quality control systems. Existing VLM-based anomaly reasoning methods are capable of detecting open-ended anomalies in general domains. However, their performance declines in industrial settings characterized by intricate object transformations, strict physics, and procedural constraints. To tackle the complexity of such interaction-intensive detection, we introduce a training-free agentic framework for anomaly detection free of domain-specific knowledge, emphasizing object state evolution like humans inspectors. It is designed to track spatial-temporal dynamics and underlying transformations of detected objects over time, and then reason over the object-wise temporal state trajectories to identify abnormal objects in grounded frames. Our method overcomes limitations of prior approaches that rely on retraining on normal clips or injecting domain knowledge as context for test-time inference. Extensive experiments on three IVAD datasets demonstrate that our method outperforms frontier VLMs, agentic frameworks, and traditional VAD methods fine-tuned on the respective datasets, while providing interpretable reports over anomaly processes and types.

Figures

Figures reproduced from arXiv: 2607.18142 by Lei Wang, Mei Yuan, Min Xu, Qifeng Wu, Qi Long, Yang Liu, Yizhou Zhao, Zhenyang Li.

Figure 1
Figure 1. Figure 1: Why object state evolution matters for industrial anomaly detec [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of O-VAD. (a) Stage 1 samples multiple frames and queries a VLM to discover all objects, then segments them with SAM to produce initial masks and metadata. (b) Stage 2 constructs spatiotemporal tubelets via CropFormer [29] and SAM2 [31], recovers missing tracks through semantic and proximity priors, and query the VLM for open-ended state change detection and inter-object interaction analysis. (c) … view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative results. Each row shows a different industrial manipulation task— plastic bottle rotation (top), hinge screw fastening (bottom)—with GPT-5’s output (left) and O-VAD’s output (right). Open-ended anomaly explanation. O-VAD can detect free-form anomalies grounded in tracked object states rather than a fixed taxonomy. In Case 1, it attributes the leakage to progressive stress from repeated clamp-in… view at source ↗
Figure 4
Figure 4. Figure 4: Multi-instance re-ID on LiquidAD. O-VAD’s iterative detect–segment– mask-out scheme assigns one stable tubelet per pipette across all 8 visually-identical instances, including under partial occlusion by the dispensing head and during liquid transfer. C.3 Comparative Reasoning Traces We compare the full reasoning traces of O-VAD against baseline VLMs on two representative success cases from LiquidAD and Phy… view at source ↗
Figure 5
Figure 5. Figure 5: Comparative reasoning traces on LiquidAD and Phys-AD. [PITH_FULL_IMAGE:figures/full_fig_p043_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Failure cases on servo and magnet categories. [PITH_FULL_IMAGE:figures/full_fig_p044_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Failure cases on button and clip categories. [PITH_FULL_IMAGE:figures/full_fig_p045_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 12 linked inside Pith

  1. [1]

    arXiv preprint arXiv:2511.21631 (2025)

    Bai, S., Cai, Y., Chen, R., Chen, K., Chen, X., Cheng, Z., Deng, L., Ding, W., Gao, C., Ge, C., et al.: Qwen3-vl technical report. arXiv preprint arXiv:2511.21631 (2025)

    Pith/arXiv arXiv 2025

  2. [2]

    arXiv preprint arXiv:2511.16719 (2025)

    Carion, N., Gustafson, L., Hu, Y.T., Debnath, S., Hu, R., Suris, D., Ryali, C., Alwala,K.V.,Khedr,H.,Huang,A.,etal.:Sam3:Segmentanythingwithconcepts. arXiv preprint arXiv:2511.16719 (2025)

    Pith/arXiv arXiv 2025

  3. [3]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Chen, Z., Li, J., Liang, J., Tan, L., Guo, Y., Lu, C., Li, Y.L.: Mˆ 3-vos: Multi-phase, multi-transition, and multi-scenery video object segmentation. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 29193–29202 (2025)

    2025

  4. [4]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Cheng, H.K., Oh, S.W., Price, B., Lee, J.Y., Schwing, A.: Putting the object back into video object segmentation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 3151–3161 (2024)

    2024

  5. [5]

    Expert Systems with Applications271, 126581 (2025)

    Dabouei, A., Shibu, J.P., Dalal, V., Cao, C., MacWilliams, A., Kangas, J., Xu, M.: Deep video anomaly detection in automated laboratory setting. Expert Systems with Applications271, 126581 (2025)

    2025

  6. [6]

    In: International conference on pattern recognition

    Defard,T.,Setkov,A.,Loesch,A.,Audigier,R.:Padim:apatchdistributionmodel- ing framework for anomaly detection and localization. In: International conference on pattern recognition. pp. 475–489. Springer (2021)

    2021

  7. [7]

    In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers)

    Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: Bert: Pre-training of deep bidi- rectional transformers for language understanding. In: Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers). pp. 4171–4186 (2019)

    2019

  8. [8]

    arXiv preprint arXiv:2412.04769 (2024)

    Fan, L., Huang, J., Di, D., Su, A., Pagnucco, M., Song, Y.: Revitalizing reconstruc- tion models for multi-class anomaly detection via class-aware contrastive learning. arXiv preprint arXiv:2412.04769 (2024)

    Pith/arXiv arXiv 2024

  9. [9]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Fang, Z., Wang, X., Li, H., Liu, J., Hu, Q., Xiao, J.: Fastrecon: Few-shot in- dustrial anomaly detection via fast feature reconstruction. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 17481–17490 (2023)

    2023

  10. [10]

    In: Proceedings of the AAAI conference on artificial intelligence

    Gu, Z., Zhu, B., Zhu, G., Chen, Y., Tang, M., Wang, J.: Anomalygpt: Detecting industrial anomalies using large vision-language models. In: Proceedings of the AAAI conference on artificial intelligence. vol. 38, pp. 1932–1940 (2024)

    1932

  11. [11]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Guo, J., Lu, S., Zhang, W., Chen, F., Li, H., Liao, H.: Dinomaly: The less is more philosophy in multi-class unsupervised anomaly detection. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 20405–20415 (2025)

    2025

  12. [12]

    Advances in Neural Information Processing Systems37, 71162– 71187 (2024)

    He, H., Bai, Y., Zhang, J., He, Q., Chen, H., Gan, Z., Wang, C., Li, X., Tian, G., Xie, L.: Mambaad: Exploring state space models for multi-class unsupervised anomaly detection. Advances in Neural Information Processing Systems37, 71162– 71187 (2024)

    2024

  13. [13]

    In: Proceedings of the 30th ACM international con- ference on multimedia

    Huang, C., Liu, Y., Zhang, Z., Liu, C., Wen, J., Xu, Y., Wang, Y.: Hierarchi- cal graph embedded pose regularity learning via spatio-temporal transformer for abnormal behavior detection. In: Proceedings of the 30th ACM international con- ference on multimedia. pp. 307–315 (2022)

    2022

  14. [14]

    arXiv preprint arXiv:2505.19877 (2025)

    Huang, C., Wang, B., Wen, J., Liu, C., Wang, W., Shen, L., Cao, X.: Vad-r1: To- wards video anomaly reasoning via perception-to-cognition chain-of-thought. arXiv preprint arXiv:2505.19877 (2025)

    Pith/arXiv arXiv 2025

  15. [15]

    In: Pro- O-VAD 17 ceedings of the IEEE/CVF winter conference on applications of computer vision

    Hyun, J., Kim, S., Jeon, G., Kim, S.H., Bae, K., Kang, B.J.: Reconpatch: Con- trastive patch representation learning for industrial anomaly detection. In: Pro- O-VAD 17 ceedings of the IEEE/CVF winter conference on applications of computer vision. pp. 2052–2061 (2024)

    2052

  16. [16]

    arXiv preprint arXiv:2410.09453 (2024)

    Jiang, X., Li, J., Deng, H., Liu, Y., Gao, B.B., Zhou, Y., Li, J., Wang, C., Zheng, F.: Mmad: A comprehensive benchmark for multimodal large language models in industrial anomaly detection. arXiv preprint arXiv:2410.09453 (2024)

    Pith/arXiv arXiv 2024

  17. [17]

    Advances in Neural In- formation Processing Systems35, 15433–15445 (2022)

    Jiang, X., Liu, J., Wang, J., Nie, Q., Wu, K., Liu, Y., Wang, C., Zheng, F.: Soft- patch: Unsupervised anomaly detection with noisy data. Advances in Neural In- formation Processing Systems35, 15433–15445 (2022)

    2022

  18. [18]

    In: The Fourteenth International Con- ference on Learning Representations (2026)

    Kang, H., Lee, W., Kim, J., Park, H.: Judo: A juxtaposed domain-oriented multi- modal reasoner for industrial anomaly qa. In: The Fourteenth International Con- ference on Learning Representations (2026)

    2026

  19. [19]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 4015–4026 (2023)

    2023

  20. [20]

    In: European Conference on Computer Vision

    Li, G., Cai, G., Zeng, X., Zhao, R.: Scale-aware spatio-temporal relation learning for video anomaly detection. In: European Conference on Computer Vision. pp. 333–350. Springer (2022)

    2022

  21. [21]

    arXiv preprint arXiv:2506.02535 (2025)

    Li, J., Dang, L., Xiao, Q., Shang, S., Cheng, J., Wu, H., Hao, Y., Wu, Q.: Video anomaly detection with semantics-aware information bottleneck. arXiv preprint arXiv:2506.02535 (2025)

    arXiv 2025

  22. [22]

    Advances in Neu- ral Information Processing Systems38, 148372–148404 (2026)

    Li, W., Xu, Y., Rao, Y., Wang, Z., Deng, S.: Vadtree: Explainable training-free video anomaly detection via hierarchical granularity-aware tree. Advances in Neu- ral Information Processing Systems38, 148372–148404 (2026)

    2026

  23. [23]

    In: Proceedings of the Computer Vision and Pattern Recogni- tion Conference

    Li, W., Gu, Y., Chen, X., Xu, X., Hu, M., Huang, X., Wu, Y.: Towards visual discrimination and reasoning of real-world physical dynamics: Physics-grounded anomaly detection. In: Proceedings of the Computer Vision and Pattern Recogni- tion Conference. pp. 30409–30419 (2025)

    2025

  24. [24]

    arXiv preprint arXiv:2511.00962 (2025)

    Lin, D., Qu, M., Han, K., Jiao, J., Jin, X., Wei, Y.: A unified reasoning frame- work for holistic zero-shot video anomaly analysis. arXiv preprint arXiv:2511.00962 (2025)

    arXiv 2025

  25. [25]

    In: 2025 International Conference on Advanced Robotics and Mechatronics (ICARM)

    Lin, S., Wang, C., Ding, X., Wang, Y., Du, B., Song, L., Wang, C., Liu, H.: A vlm- based method for visual anomaly detection in robotic scientific laboratories. In: 2025 International Conference on Advanced Robotics and Mechatronics (ICARM). pp. 34–39. IEEE (2025)

    2025

  26. [26]

    IEEE Transactions on Circuits and Systems for Video Technology35(1), 380–393 (2024)

    Liu, J., Yan, Y., Li, J., Zhao, W., Chu, P., Sheng, X., Liu, Y., Yang, X.: Ipad: Industrial process anomaly detection dataset. IEEE Transactions on Circuits and Systems for Video Technology35(1), 380–393 (2024)

    2024

  27. [27]

    In: Proceedings of the 27th ACM International Conference on Multimedia

    Liu, K., Ma, H.: Exploring background-bias for anomaly detection in surveillance videos. In: Proceedings of the 27th ACM International Conference on Multimedia. pp. 1490–1499 (2019)

    2019

  28. [28]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Park, H., Noh, J., Ham, B.: Learning memory-guided normality for anomaly detec- tion. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14372–14381 (2020)

    2020

  29. [29]

    arXiv preprint arXiv:2211.05776 (2022)

    Qi, L., Kuen, J., Guo, W., Shen, T., Gu, J., Jia, J., Lin, Z., Yang, M.H.: High- quality entity segmentation. arXiv preprint arXiv:2211.05776 (2022)

    Pith/arXiv arXiv 2022

  30. [30]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PmLR (2021) 18 M. Yuan et al

    2021

  31. [31]

    arXiv preprint arXiv:2408.00714 (2024)

    Ravi, N., Gabeur, V., Hu, Y.T., Hu, R., Ryali, C., Ma, T., Khedr, H., Rädle, R., Rolland, C., Gustafson, L., et al.: Sam 2: Segment anything in images and videos. arXiv preprint arXiv:2408.00714 (2024)

    Pith/arXiv arXiv 2024

  32. [32]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Roth, K., Pemula, L., Zepeda, J., Schölkopf, B., Brox, T., Gehler, P.: Towards total recall in industrial anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 14318–14328 (2022)

    2022

  33. [33]

    arXiv preprint arXiv:2601.03267 (2025)

    Singh, A., Fry, A., Perelman, A., Tart, A., Ganesh, A., El-Kishky, A., McLaughlin, A., Low, A., Ostrow, A., Ananthram, A., et al.: Openai gpt-5 system card. arXiv preprint arXiv:2601.03267 (2025)

    Pith/arXiv arXiv 2025

  34. [34]

    In: 2023 IEEE International Conference on Multimedia and Expo (ICME)

    Sun, S., Gong, X.: Long-short temporal co-teaching for weakly supervised video anomaly detection. In: 2023 IEEE International Conference on Multimedia and Expo (ICME). pp. 2711–2716. IEEE (2023)

    2023

  35. [35]

    arXiv preprint arXiv:2511.04678 (2025)

    Sun, Y., Yang, X., Sun, J.J., Hariharan, B.: Tracking and understanding object transformations. arXiv preprint arXiv:2511.04678 (2025)

    arXiv 2025

  36. [36]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Tokmakov, P., Li, J., Gaidon, A.: Breaking the" object" in video object segmen- tation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22836–22845 (2023)

    2023

  37. [37]

    In: European Conference on Computer Vision

    Wu, J.C., Hsieh, H.Y., Chen, D.J., Fuh, C.S., Liu, T.L.: Self-supervised sparse representation for video anomaly detection. In: European Conference on Computer Vision. pp. 729–745. Springer (2022)

    2022

  38. [38]

    arXiv preprint arXiv:2108.03825 (2021)

    Wu, J., Zhang, W., Li, G., Wu, W., Tan, X., Li, Y., Ding, E., Lin, L.: Weakly- supervised spatio-temporal anomaly detection in surveillance video. arXiv preprint arXiv:2108.03825 (2021)

    Pith/arXiv arXiv 2021

  39. [39]

    In: Proceedings of the 32nd ACM International Conference on Multimedia

    Wu, P., Zhou, X., Pang, G., Yang, Z., Yan, Q., Wang, P., Zhang, Y.: Weakly supervisedvideoanomalydetectionandlocalizationwithspatio-temporalprompts. In: Proceedings of the 32nd ACM International Conference on Multimedia. pp. 9301–9310 (2024)

    2024

  40. [40]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Ye, M., Liu, W., He, P.: Vera: Explainable video anomaly detection via verbalized learning of vision-language models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 8679–8688 (2025)

    2025

  41. [41]

    Yu, J., Li, X., Zhao, X., Zhang, H., Wang, Y.X.: Video state-changing object seg- mentation.In:ProceedingsoftheIEEE/CVFinternationalconferenceoncomputer vision. pp. 20439–20448 (2023)

    2023

  42. [42]

    In: Proceedings of the IEEE/CVF international conference on computer vision

    Zavrtanik, V., Kristan, M., Skočaj, D.: Draem-a discriminatively trained re- construction embedding for surface anomaly detection. In: Proceedings of the IEEE/CVF international conference on computer vision. pp. 8330–8339 (2021)

    2021

  43. [43]

    Zhang, H., Wang, Z., Wu, Z., Jiang, Y.: Diffusionad: norm-guided one-step denois- ing diffusion for anomaly detection (2023)

    2023

  44. [44]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhang, X., Xu, M., Zhou, X.: Realnet: A feature selection network with realis- tic synthetic anomaly for anomaly detection. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16699–16708 (2024)

    2024

  45. [45]

    arXiv preprint arXiv:2505.22039 (2025)

    Zhao, S., Lin, Y., Han, L., Zhao, Y., Wei, Y.: Omniad: Detect and understand industrial anomaly via multimodal reasoning. arXiv preprint arXiv:2505.22039 (2025)

    Pith/arXiv arXiv 2025

  46. [46]

    arXiv preprint arXiv:2604.12440 (2026)

    Zheng, H., Lin, T., Wang, W., Wang, Z., Zhang, W., Zhu, J., Shao, F.: Iad-unify: A region-grounded unified model for industrial anomaly segmentation, understand- ing, and generation. arXiv preprint arXiv:2604.12440 (2026)

    Pith/arXiv arXiv 2026

  47. [47]

    Advances in neural information processing systems36, 46595–46623 (2023) O-VAD 19

    Zheng, L., Chiang, W.L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E., et al.: Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in neural information processing systems36, 46595–46623 (2023) O-VAD 19

    2023

  48. [48]

    arXiv preprint arXiv:2505.23504 (2025)

    Zhu, L., Chen, Q., Shen, X., Cun, X.: Vau-r1: Advancing video anomaly under- standing via reinforcement fine-tuning. arXiv preprint arXiv:2505.23504 (2025)

    Pith/arXiv arXiv 2025

  49. [49]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Zou, S., Tian, X., Wesemann, L., Waschkowski, F., Yang, Z., Zhang, J.: Unlocking vision-language models for video anomaly detection via fine-grained prompting. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 4223–4233 (2026) 20 M. Yuan et al. Appendix A More Details about Experiment Settings A.1 Details of Dataset...

    2026

  50. [50]

    Grounding Uniform scan + max-confidence selection 5 frames for SAM prompt; all frames retained deterministic

  51. [51]

    Tracking Inter-stage VLM, content-adaptive FPS FPS∈[2,10](default 3 if VLM unavailable) content-adaptive

  52. [52]

    screw” and “screw thread

    Reasoning Dataset-specific stride for verification 60/30/25 (Phys-AD/LiquidAD/IPAD) deterministic 24 M. Yuan et al. Object deduplication (whole vs. parts).Stage 1 issues asingleopen-vocabulary VLM query requesting all distinct objects, and SAM is invoked with the full name list as a multi-class query that producesone fused binary mask. Whole and part labe...

  53. [53]

    Detecting OBVIOUS change in object appearance, shape, material or moving state (slight changes should be ignored)

  54. [54]

    Recognizing interactions between objects

  55. [55]

    "role":

    Check for dynamics like material flow, deformation, or leaking." "role": "user", "content": ( f"{ ˜I i tk}" f"{ ˜I i tk+1}" "Compare these two frames from a video." "The same object is highlighted with {init_c_name} contour in both." "First image: Earlier frame Second image: Later frame" "Carefully check for OBVIOUS changes:" "- DEFORMATION: Is the object...

    arXiv 2001

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.