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

01 / Abstract

Reasoning over object state evolution, like a human inspector

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 can detect open‑ended anomalies in general domains, but 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 O‑VAD, a training‑free agentic framework free of domain‑specific knowledge that emphasizes object state evolution. It tracks the spatial‑temporal dynamics and underlying transformations of detected objects over time, then reasons over the object‑wise temporal state trajectories to identify abnormal objects in grounded frames — overcoming prior approaches that rely on retraining on normal clips or injecting domain knowledge at test time.

Extensive experiments on three IVAD datasets show O‑VAD outperforms frontier VLMs, agentic frameworks, and traditional VAD methods fine‑tuned on the respective datasets, while providing interpretable reports over anomaly processes and types.

02 / Method

A ground → track → reason pipeline

Given an industrial video, O‑VAD produces a structured anomaly report — type, severity, affected object, temporal localization, and a natural‑language causal explanation — without any domain‑specific knowledge, predefined taxonomy, or training data. It runs in three agentic stages.

03 / Quantitative results

State of the art without any training

We foreground the metrics that matter for open-ended IVAD — detection AUROC (video and frame level) and anomaly-type description quality — and drop Acc/P/R/F1. O‑VAD posts the best video-level AUROC on all three datasets and the most faithful type descriptions, then leads or ties across most Phys-AD and IPAD categories.

04 / Qualitative results

Grounded reports where frontier VLMs miss the anomaly evidence.

We compare O‑VAD against GPT‑5 prompted with the same six‑step chain‑of‑thought but without object grounding or state tracking. Lacking object‑wise evidence, GPT‑5 defaults to “no anomaly”; O‑VAD localizes the object, frame range, severity, and root cause.

4.1 / CASE COMPARISONS

Servo motorSpecification-dependent

A restricted rotation angle is labeled normal by both O‑VAD and the VLM baseline — the small‑angle motion is visually indistinguishable from a normal commanded rotation.

⤢ Full size

Degaussed magnetInvisible property

O‑VAD detects the over‑clamping deformation as the magnet drops, but misattributes the cause: a demagnetized magnet looks identical to an ordinary red plastic block, so loss of magnetism is unobservable.

⤢ Full size

Unpressable clipNo observable evidence

A clip that cannot be pressed produces no observable state change across frames, so both methods default to a normal prediction.

⤢ Full size

Challenging categories. O‑VAD’s weakest Phys‑AD AUROC lands on Caster Wheel (0.301), Zipper (0.388), Button (0.393), and Slide (0.414) — subtle rotational resistance, small localized transitions missed at the default sampling interval, and alignment hard to judge from cropped views. These motivate adaptive temporal sampling and multi‑scale spatial reasoning.

Mean ratings (1–5, higher better) unless marked %; O‑VAD in accent with a fill bar; “–” = URF-ZS-HVAA near-floor value not itemized in the paper. Q3 and Q5 human per-method scores are reported only as part-level aggregates.

06 / DEMO

Method walkthrough

A short overview of the ground → track → reason pipeline and representative anomaly reports.

@inproceedings{yuan2026ovad,
  title     = {O-VAD: Industrial Video Anomaly Detection
               through Object-Centric Tracking and Reasoning},
  author    = {Yuan, Mei and Long, Qi and Wu, Qifeng and Li, Zhenyang and
               Zhao, Yizhou and Wang, Lei and Liu, Yang and Xu, Min},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}