Automated Anomaly Detection in Industrial Videos

Summary & Methodology Analysis

The O-VAD architecture operates through a three-stage pipeline designed to handle complex industrial scenes. In the first stage, the system employs VLM-grounded masking using SAM3 to perform automated object discovery and segmentation. The second stage builds spatiotemporal tubelets by leveraging CropFormer for entity segmentation and SAM2 for temporal propagation. These tracks are recovered using spatial proximity and semantic consistency priors, followed by open-ended state change detection and interaction analysis via VLM querying at temporal transitions. This sequence enables the system to isolate individual objects and their interactions within a complex video stream.

For the final stage of analysis, O-VAD utilizes a cascaded chain-of-thought (CoT) reasoning module, which is a technique that prompts the model to break down complex problems into intermediate logical steps to improve output accuracy. This is paired with multiplicative confidence gating to visually verify candidate anomalies against caption and frame evidence. By accumulating evidence through this multi-stage reasoning, the system can determine if a detected change constitutes a true anomaly based on visual and semantic context.

The system faces notable practical trade-offs. The multi-stage pipeline introduces higher latency compared to single-forward-pass models, and this latency scales linearly with the number of tracked objects. Furthermore, the system struggles with anomalies defined by invisible physical properties or specifications that cannot be inferred visually. Finally, static or perception-ambiguous defects that exhibit visually plausible behavior remain a fundamental challenge for the current architecture.