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

Source-linked AI summary

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

arXiv:2607.18142v1Jul 20, 2026cs.CVcs.AIcs.CLcs.MA

TL;DR

Industrial video anomaly detection challenges general-purpose VLMs because industrial processes involve complex object transformations and strict physical and procedural constraints. O-VAD is a training-free agentic framework that tracks object state evolution and reasons over object-wise trajectories, achieving state-of-the-art performance across three IVAD datasets while producing interpretable anomaly reasoning.

Problem
Existing VLM methods struggle to reliably reason about localized object-centric state changes and interactions over time in industrial processes.
Method
O-VAD uses a training-free ground–track–reason pipeline to discover objects, construct spatiotemporal trajectories, and perform chain-of-thought anomaly reasoning without domain-specific knowledge or fine-tuning.
Results
O-VAD achieves state-of-the-art anomaly detection across three IVAD datasets at both video and frame levels, outperforming frontier VLMs, agentic frameworks, and fine-tuned traditional VAD methods.
Takeaways & Limitations
The framework provides grounded, interpretable reasoning over anomaly processes and open-ended anomaly types without labor-intensive annotations.
Takeaways & Limitations
O-VAD has higher latency than single-pass VAD models and performs worse on anomalies requiring invisible physical properties, precise specifications, or visually ambiguous defects.

Abstract

from arXiv · show

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.

1 Introduction

O-VAD addresses the challenges of industrial video anomaly detection by tracking object state evolution in a training-free, domain-knowledge-free agentic framework. It reports state-of-the-art video- and frame-level performance on three IVAD datasets with interpretable anomaly reasoning.

Motivation: IVAD targets anomalous objects and events in industrial processes, supporting automated inspection for reduced costs and improved production reliability.Industrial processes are challenging because objects undergo complex transformations.
Limitations: Traditional reconstruction- and embedding-based VAD methods typically require plentiful normal samples for each object class under a one-class-one-model paradigm.These methods detect anomalies through reconstruction error or deviations from modeled normal embeddings.
Limitations: VLM performance declines in industrial settings because of limited domain-specific knowledge and fine-grained annotations, while external context and complex prompts can misalign responses.Video- and frame-level features also poorly capture localized object details over time for interaction-intensive reasoning.
Method: O-VAD uses a training-free “ground→track→reason” pipeline to discover objects, construct spatiotemporal tubelets, detect open-ended state changes, and verify anomaly reasoning.The framework requires no domain-specific knowledge, labels, predefined anomaly taxonomy, or fine-tuning.
Results and contributions: O-VAD achieves state-of-the-art video- and frame-level performance on IPAD, Phys-AD, and LiquidAD, surpassing frontier VLMs, agentic VAD frameworks, and fine-tuned traditional methods.It provides interpretable reasoning over anomaly processes and open-ended types without labor-intensive annotations.

2 Related Work

Related work spans traditional video anomaly detection, which models reconstruction errors or deviations from normal feature distributions, and newer VLM-based methods that emphasize generalizability and explainability in zero- and few-shot settings. VLM approaches commonly use prompt-based video captioning and anomaly scoring, sometimes augmented with temporal modeling or prompting.

Traditional Video Anomaly Detection: Traditional methods flag anomalies through elevated reconstruction error or deviations from modeled normal-sample feature distributions.These correspond to reconstruction-based and embedding-based paradigms, respectively.
Traditional Video Anomaly Detection: Both traditional paradigms follow a “one-class-one-model” approach that requires abundant normal samples.The passage identifies this as one of their shared limitations.
Multimodal Video Anomaly Detection: VLM-based methods are valued for generalizability and explainability, particularly in zero- and few-shot anomaly detection.These approaches revisit anomaly detection using multimodal video understanding.
Multimodal Video Anomaly Detection: These VLM approaches use sophisticated prompts to caption videos and assign anomaly scores with off-the-shelf VLMs.Some methods additionally incorporate auxiliary temporal modeling or prompting to improve spatio-temporal awareness.

3 Method

O-VAD uses a training-free, object-centric three-stage pipeline that discovers and segments objects, tracks their state trajectories and changes, then performs open-ended anomaly reasoning. Its final reasoning is grounded in accumulated object evidence and produces interpretable anomaly decisions without a fixed taxonomy.

Pipeline overview: The three stages discover task-relevant objects, track per-object state trajectories and changes, and reason over accumulated evidence with cascaded chain-of-thought.Stage 1 uses VLM-grounded masking with SAM; Stage 2 tracks objects through transformations; Stage 3 separates abnormal stages from normal process behavior.
Stage 1: Object discovery and segmentation: Sampling multiple candidate frames enables discovery of initially occluded or absent objects, whose inventories are merged and segmented with SAM.Initial masks and object metadata ground subsequent tracking and reasoning in concrete per-object evidence.
Stage 2: Tracking and state changes: Spatiotemporal tubelets track objects across frames, while new tubelets and spatial-semantic priors recover tracks after transformations such as breaking apart or releasing material.CropFormer provides per-frame partitions and SAM2 tracks entities forward; valid continuation tracks require both proximity and semantic consistency.
Stage 2: Tracking and state changes: The VLM labels state changes with open-ended types and causes, representing each event by temporal bounds, type, cause, description, severity, and affected object.Severity is categorized as none, slight, moderate, or severe.
Stage 3: Cognitive anomaly reasoning: Cascaded chain-of-thought reasoning uses object metadata, filtered state changes, task context, captions, and frames to produce freely reasoned anomaly sets with complete reasoning traces.The process identifies the industrial task, triages state changes, and cites object IDs, frame ranges, change types, and severities.

4 Experiment

Experiments on three industrial benchmarks show that training-free O-VAD outperforms competing VLM, agentic, and trained VAD methods while supporting object-centric anomaly reasoning. Per-dataset results, case analyses, and ablations demonstrate the importance of state trajectories, structured reasoning, and verification.

Datasets & evaluation metrics: O-VAD evaluates official test splits from three industrial benchmarks covering diverse anomaly types and interaction modalities, using video-, type-, and frame-level metrics.Phys-AD includes 6,434 videos, 22 object categories, and 47 defect types requiring reasoning over pressing, rotating, and grasping interactions.
Overall comparison: O-VAD achieves the best average training-free video-level AUROC on Phys-AD at 0.584, exceeding Qwen3-VL-32B at 0.513, GPT-5 at 0.503, URF-ZS-HVAA at 0.426, VERA at 0.456, MNAD.p at 0.495, and S3R at 0.555.It also obtains the best training-free type-level BERTScore at 0.803, with the same overall pattern on LiquidAD and IPAD.
Per-dataset analysis: O-VAD attains the best or second-best AUROC on 16 of 22 Phys-AD categories, including Gear at 0.879, Lock at 0.780, Screw at 0.704, Sticky Roller at 0.811, and U Disk at 0.606.The largest gains occur for complex multi-step interactions and subtle surface-level anomalies.
Per-dataset analysis: On LiquidAD, O-VAD achieves the best training-free video-level AUROC at 0.692 and the strongest frame-level recall and F1 while tracking per-pipette state trajectories in cluttered scenes.LiquidAD contains up to eight simultaneous pipettes, where zero-shot VLMs show near-zero recall and a strong bias toward “normal.”
Case Analysis: Case analyses show O-VAD detecting progressive deformation, material release, repeated screw back-out, head-slot stripping, and shaft bending that GPT-5 misses in representative Phys-AD examples.Its reasoning traces support high-severity loss-of-containment classification and coherent hypotheses about torque-control failure.
Ablation studies: Ablations show state tracking is indispensable: removing it reduces precision, recall, and F1 to zero on three of four categories, while structured reasoning improves AUC by 0.038–0.091.Removing video-level captioning causes AUC drops of 0.247 and 0.131 for sticky roller and rubber band, while removing verification degrades AUC by up to 0.178 on sticky roller.

5 Conclusion · Appendix · A More Details about Experiment Settings

O-VAD delivers state-of-the-art industrial video anomaly detection without labor-intensive annotations, surpassing frontier VLMs while supporting more reliable, explainable, and generalizable inspection. Its multi-stage design increases latency and struggles with anomalies requiring invisible physical properties or precise specifications.

5 Conclusion: O-VAD achieves state-of-the-art performance on any-level anomaly detection without labor-intensive annotations.
5 Conclusion: The framework surpasses frontier VLMs in both quantitative and qualitative evaluations.
5 Conclusion: O-VAD establishes an industrial anomaly-detection paradigm bridging low-level pattern recognition and high-level cognitive reasoning.
5 Conclusion: This paradigm aims to enable more reliable, explainable, and generalizable industrial inspection systems.
5 Conclusion: The multi-stage agentic pipeline incurs substantially higher latency than single-forward-pass VAD models.Latency scales with the number of tracked objects per video.
5 Conclusion: O-VAD relies on the VLM’s internalized commonsense rather than domain-specific expert priors.
5 Conclusion: Performance declines on anomalies defined by invisible physical properties or precise specifications that visual appearance alone cannot reveal.The passage mentions few-shot in-context learning with specification examples as a response, but the supplied text is truncated.

A.1 Details of Datasets. … B Detailed Per-Category Quantitative Results

The supplementary sections detail O-VAD’s datasets, thresholds, sampling and prompting procedures, evaluation of open-ended anomaly types, and stage-specific model choices. Together, they connect each design decision to industrial detection challenges and report supporting ablations and evaluation protocols.

A.1 Details of Datasets.: Phys-AD, LiquidAD, and IPAD provide complementary challenges in physics-grounded interaction reasoning, per-object temporal localization, and periodic temporal structure with synthetic-to-real generalization.Phys-AD contains 6,434 videos across 47 anomaly types and 22 object categories; LiquidAD contains 2,251 videos of eight-pipette transfers; IPAD contains 597,979 frames across around 2,000 clips.
A.2 Hyperparameter Details.: O-VAD uses permissive discovery and tracking thresholds, stricter semantic filtering, confidence bypass and discard rules, and post-verification retention at τconf = 0.3.The post-verifier can improve separation between genuine and ambiguous cases, with ablation degradation of up to 0.178 AUC when removed.
A.2 Hyperparameter Details.: Removing post-verification degrades AUC by up to 0.178, while τsem = 0.7 is near-optimal and remains stable across the [0.5, 0.8] range.Performance drops noticeably only at 0.9 in the reported grid search.
A.3 Stage-wise Sampling and Grounding Details: O-VAD samples frames differently by stage: Stage 1 inventories objects once, Stage 2 uses content-adaptive sampling clamped to [2] FPS, and downstream prompts preserve object-level spatial and state information.Whole and part labels collapse into one connected mask region, while downstream trajectories retain OBJECT, STATE, MATERIAL, and BBOX fields.
A.4 Anomaly Type Evaluation.: Open-ended anomaly types are evaluated with BERT-base-uncased cosine similarity against ground-truth anomaly labels or “normal,” rather than exact-match accuracy.Candidate strings concatenate predicted anomaly_type and anomaly_subtype fields across detected anomalies.
A.5 Full prompts of proprietary models.: The prompts establish object inventories and semantic baselines, compare highlighted objects across time for state changes and interactions, then reason over accumulated evidence in a six-step anomaly chain.Stage 3 produces free-form classifications with severity, while intermediate-confidence candidates undergo visual verification against evidence frames and captions.
A.6 Segmentation and Tracking Model Selection.: O-VAD assigns SAM3 to concept-aware object segmentation, CropFormer to exhaustive class-agnostic entity segmentation, and SAM2 to temporally consistent tubelet propagation.These roles match the distinct requirements of text-to-mask, frame-to-all-masks, and mask-to-mask-sequence processing.
B Detailed Per-Category Quantitative Results: Replacing CropFormer with SAM2 automasks degrades J by 1.7 points, while replacing SAM2 with Cutie degrades J by 3.3 and temporal recall TR by 9.3.The results support specialized model selection, particularly the importance of SAM2’s tracking quality for partition completeness.

B.1 Detailed Results on Phys-AD

On Phys-AD, O-VAD achieves the strongest type-level performance among training-free methods and remains competitive with or exceeds supervised methods without training data, domain knowledge, or a predefined taxonomy. Its main weaknesses occur in categories with subtle visual cues or rapid state transitions.

Type-level results: 0.803 is O-VAD’s highest average type-level BERT score among training-free methods, with the best score on 14 of 22 categories.This supports semantically closer anomaly descriptions through object-centric state tracking.
Overall comparison: O-VAD consistently outperforms training-free alternatives and matches or sometimes exceeds fully supervised methods without training data, domain knowledge, or predefined taxonomy.Its detection strategy is aggressive, using post-verification to suppress false positives rather than abstaining under uncertainty.
Baseline coverage: 0.612 is S3R’s highest average AUROC, driven by Roll. Bear. 0.963, Sticky Roller 0.864, and Rub. Band 0.853, but it falls on Hinge 0.108, Lock 0.361, and Clip 0.587.The passage attributes this uneven coverage to strong performance on distinctive anomaly distributions but weakness on categories requiring temporal reasoning.
Failure modes: O-VAD’s weakest video-level AUROC values are Caster Wheel 0.301, Zipper 0.388, Button 0.393, and Slide 0.414.These categories challenge perception-driven reasoning because anomalies may involve subtle rotational resistance changes or small, fast state transitions.

B.2 Detailed Results on IPAD … C.2 Multi-Instance Re-Identification

On IPAD, O-VAD leads training-free methods across key video- and frame-level metrics and 12 of 16 scenarios, while revealing weaknesses on subtle, multi-object subsets. Qualitative analyses assess intermediate-state reliability and multi-instance tracking, including eight visually identical pipettes under occlusion.

B.2 Detailed Results on IPAD: O-VAD achieves the highest average video-level AUROC (56.5%) and similar frame-level AUROC (51.8%) among training-free methods.It also ranks highest in average video-level Accuracy, Revall, F1 score and average frame-level Recall, F1 score.
B.2 Detailed Results on IPAD: O-VAD secures the best video-level performance among all methods in 12 out of 16 industrial scenarios.The IPAD comparison covers video-level and frame-level AUROC across 16 industrial scenarios.
B.2 Detailed Results on IPAD: O-VAD is weakest on S05 (39.6%), S08 (38.9%), R02 (43.3%) and R03 (46.7%) video-level AUROC.S05 relies on minimal cues for object location and orientation, while S08 contains multiple main objects and requires reasoning about normal rules.
C More Qualitative Results.: The qualitative results examine O-VAD’s object-centric reasoning across industrial scenarios and its intermediate outputs.The supplied passages specify reliability and multi-instance re-identification analyses rather than additional aggregate results for this subsection.
C.1 Reliability of Intermediate Outputs: Stage-2 reliability is assessed without ground-truth state-change annotations using count, consistency and temporal-coherence axes.Table 10 reports source duration and FPS alongside Stage-1 grounded-object and Stage-2 detected-state counts, summarized by minimum, average and maximum.
C.1 Reliability of Intermediate Outputs: The post-verifier’s three thresholds define a skip/verify/discard confidence partition with one effective hyperparameter: the verification band.Removing the post-verifier causes AUROC drops up to 0.178 on sticky roller, while a ±0.1 threshold sweep leaves the retained result stable.
C.2 Multi-Instance Re-Identification: O-VAD re-identifies look-alike instances by iteratively detecting, segmenting and masking each instance until no further instance is returned.On LiquidAD, the scheme assigns one stable tubelet per pipette across all 8 visually identical instances, including partial occlusion and liquid transfer.

C.3 Comparative Reasoning Traces

Comparative traces show that baseline VLMs can produce fluent normality judgments while missing localized, object-specific anomalies. O-VAD instead uses per-object state trajectories to expose subtle events and support cross-object causal diagnosis.

LiquidAD: unintended single-channel drip: In LiquidAD, Qwen3-VL-32B reports normal pipette operation and misses a single-tip drip at frames 30–40.The drip affects only one channel among many, making it difficult for holistic frame-level inspection to detect.
LiquidAD: unintended single-channel drip: O-VAD tracks the pipette head with per-frame state annotations and flags material release from one tip at frames 30–40.The flagged tip is the second from the right.
Phys-AD: sticky roller crushing: In Phys-AD, GPT-5 describes a robotic gripper’s roller rotation and stopping as normal controlled manipulation and reports no anomaly.Its reasoning states that no deviations can be established against expectations.
Phys-AD: sticky roller crushing: O-VAD independently tracks the cardboard core, metal gripper component, and paper towel roll in the Phys-AD case.The passage identifies these as obj_id 1, obj_id 2, and a third tracked object, respectively.
Comparative conclusion: Together, the cases show that baseline VLMs lack fine-grained object-level evidence, whereas O-VAD’s state trajectories support subtle anomaly detection and multi-object causal diagnosis.The structured trajectories provide an evidential layer linking object-state changes across time and objects.

C.4 Failure Cases Analysis

The failure cases expose two limitations of perception-driven anomaly reasoning: specification-dependent anomalies require invisible quantitative thresholds, while perception-ambiguous anomalies produce visually plausible behavior. O-VAD tracks observable symptoms but cannot infer hidden physical properties or functional requirements from appearance alone.

Failure case categories: The four representative failures comprise specification-dependent anomalies and perception-ambiguous anomalies.Specification-dependent cases rely on quantitative thresholds absent from visual evidence; perception-ambiguous cases resemble normal behavior visually.
Specification-dependent failures: In the servo angle-restricted case, O-VAD and Qwen3-VL-32B classify smooth limited rotation as normal because the expected angular range is visually unavailable.Neither state tracking nor commonsense reasoning can identify the deviation without the expected rotation angle.
Specification-dependent failures: In the unable-to-press case, O-VAD detects no clip state change because the pressing tool moves while the clip remains static.The passage presents this as another specification-dependent failure involving a functional outcome not recoverable from the observed motion alone.
Perception-ambiguous failures: In the button-stuck case, O-VAD flags tool or fixture damage from repeated deformation, although the actual anomaly is the button’s failure to actuate.A stuck button and a normally stiff button produce similar visual signatures of tool compression.
Perception-ambiguous failures: In the magnet degaussing case, O-VAD attributes detachment to crushing and dropping from over-clamping, but cannot infer the magnet’s invisible loss of adhesion.The magnet visually resembles an ordinary red plastic block, preventing either method from recovering its magnetic properties.
Common limitation: Across these cases, O-VAD captures observable symptoms but cannot bridge perceptual evidence to hidden specifications, physical properties, or required object functions.The cited hidden factors include magnetic strength, required actuation force, angular specification, and knowledge of intended object behavior.

C.5 Human Evaluation

A human study with five domain experts evaluated anonymized reports from O-VAD, Qwen3-VL-32B, and URF-ZS-HVAA on 10 Phys-AD videos using Likert-scale ratings and pairwise preferences. O-VAD ranked first in 82% of cases and achieved the highest scores across all nine evaluated dimensions, including detection, explanation quality, and actionability.

Human Evaluation Protocol: Five domain experts rated anonymized reports from three methods on 10 Phys-AD videos across nine Likert-scale dimensions and pairwise preference ranking.The evaluation used 7 anomalous and 3 normal videos.
Overall comparison: 82%: O-VAD was ranked first in 41 of 50 evaluator–video pairs and achieved the highest scores on all nine dimensions.Its average detection score was 4.42, compared with 2.89 for Qwen3-VL-32B and 1.14 for URF-ZS-HVAA.
Detection correctness: 4.98: O-VAD’s binary detection score exceeded Qwen3-VL-32B’s 2.70 and URF-ZS-HVAA’s 1.24.Evaluators found near-perfect agreement that O-VAD correctly identified anomalous versus normal videos.
Detection correctness: O-VAD led anomaly type identification at 4.26 vs. 2.60 vs. 1.08 and temporal localization at 3.96 vs. 2.32 vs. 1.08.The finer-grained dimensions showed more variance because open-ended type classification and frame-level precision are difficult.
Explanation quality and actionability: 4.58: O-VAD’s completeness score exceeded 3.36 for Qwen3-VL-32B and 1.00 for URF-ZS-HVAA, while causal coherence reached 4.30 vs. 3.02 vs. 1.00.The results associate O-VAD’s advantage with state-trajectory-grounded reports and structured chain-of-thought reasoning.
Anomalous vs. normal cases: 5.00: O-VAD achieved perfect binary correctness on the three normal videos, while both baselines scored below 1.80 and Qwen3-VL-32B fell to 1.13.This indicates severe false-positive tendencies, particularly for Qwen3-VL-32B, whose score was 3.37 on anomalous videos.

C.6 LLM-as-Judge

Using GPT-4o as a structured LLM judge on 10 Phys-AD videos, O-VAD ranks first overall and leads detection correctness, while Qwen3-VL-32B scores higher on explanation quality. Anomaly type correctness remains difficult for both methods, and the results broadly corroborate human evaluation trends.

Evaluation protocol: The protocol uses GPT-4o with structured prompts to mirror the human questionnaire and scale evaluation beyond the 10-video human study.It is applied across the same 10 Phys-AD videos; Q4 temporal localization is excluded because the judging set lacks frame-level ground-truth annotations.
Overall comparison: O-VAD is ranked first in 70% of cases, ahead of Qwen3-VL-32B at 30% and URF-ZS-HVAA at 10%.O-VAD also achieves the highest overall quality, Q9: 3.50 vs. 2.70 vs. 1.20, and diagnostic usefulness, Q8: 3.40 vs. 2.30 vs. 1.10.
Detection vs. explanation: O-VAD leads detection with avg. 3.20 vs. 2.33 vs. 1.13 and binary correctness, Q1: 4.60 vs. 2.60 vs. 1.40, but Qwen3-VL-32B leads explanation scores.Qwen3-VL-32B scores higher on faithfulness, 4.90 vs. 3.90, completeness, 4.00 vs. 3.90, and causal coherence, 3.70 vs. 3.60.
Anomaly type correctness: Both O-VAD and Qwen3-VL-32B tie on type correctness, Q2: 1.80, their lowest-scoring dimension.The passage attributes this difficulty to open-ended anomaly type classification and precise matching of the ground-truth anomaly taxonomy.
Consistency with human evaluation: The LLM-as-judge results broadly corroborate human evaluation: O-VAD ranks first overall, URF-ZS-HVAA is weakest, and O-VAD’s advantage centers on detection correctness and actionability.The main discrepancy is explanation quality, where the judge favors Qwen3-VL-32B’s fluent but often factually incorrect narratives.

D Algorithm of O-VAD

O-VAD is a training-free, domain-agnostic pipeline that detects industrial anomalies by tracking object state evolution and reasoning over temporal trajectories. It combines automated object grounding, object-centric state-change detection, and state-aware reasoning with confidence-tiered verification.

Stage 1: Automated Object Grounding: Stage 1 discovers and segments objects using VLM prompts and SAM-based grounding.Candidate frames yield object inventories containing names, descriptions, and spatial cues; objects are merged and deduplicated before segmentation.
Stage 2: Object-Centric State Tracking: Stage 2 constructs tubelets, recovers missing tracks, identifies initial states, and queries sampled frame pairs for open-ended object state changes.Each detected change records its frame interval, type, cause, description, severity, and object identity.
Stage 3: State-Aware Anomaly Reasoning: Stage 3 performs six-step anomaly reasoning—observe, expect, compare, cause, classify, and assess severity—using sampled key frames and tracked evidence.The reasoning pass produces raw anomalies for subsequent confidence-based handling.
Stage 3: State-Aware Anomaly Reasoning: Confidence-tiered visual verification retains high-confidence anomalies, discards low-confidence anomalies, and verifies intermediate cases before adjusting final confidence.Verification compares anomaly claims against captions and key frames, multiplying original confidence by verification confidence or its complement.
Design Principle: All three stages are training-free and require neither a predefined anomaly taxonomy nor domain-specific knowledge.The framework is designed to reason over object-wise temporal state trajectories without retraining on normal clips or injecting domain knowledge at test time.