TL;DR
VerdictAD is a training-free, plug-and-play framework that uses "typed visual decisions" to refine the output of existing anomaly detectors. By acting as a decision-guided sanity check, it effectively filters out false positives caused by normal variations and significantly improves the detection of complex logical anomalies in few-shot industrial inspection.
What Problem Does This Paper Address?
Industrial anomaly detection (AD) is moving toward the "few-shot" regime, where a system must learn to identify defects using only a handful of normal reference images. Most current methods rely on measuring feature differences between a query image and these references.
However, this approach faces two major hurdles:
- False Positives: Normal variations—such as a change in the color of juice in a bottle or the orientation of a screw—can create large feature differences that detectors mistake for defects.
- Logical Anomalies: Some defects are "logical" rather than structural. For example, a breakfast box missing a piece of dried fruit might not show a strong local feature difference, but it is still a defective product.
The authors argue that we shouldn't just focus on better visual representations; we need a way to verify if a detected difference is actually a defect.
Methods
VerdictAD introduces a training-free refinement layer that can be added to almost any existing anomaly detector. It leverages Typed Decision Models (like Clef-Flash), which are vision-language models designed to output probabilities for specific questions (e.g., "Is this anomalous?") rather than generating free-form text.
The framework consists of two main modules:
- Global Verdict Refinement (GVR): This refines the overall image-level anomaly score. It asks the decision model two questions: one about the query image alone (reference-free) and one comparing the query to a grid of normal references (reference-conditioned). The resulting probability acts as a multiplicative gate, suppressing scores that the decision model deems "normal."
- Local Verdict Refinement (LVR): This refines the pixel-level anomaly map. It uses a sliding window to evaluate local regions through two views:
- Crop Decision: A zoomed-in view to catch tiny details like cracks or scratches.
- Box Decision: A view of the full image with a red box around the region of interest. This preserves global context, which is essential for identifying logical anomalies (like a missing component in a specific spot).
Datasets
The authors evaluated VerdictAD on three major industrial benchmarks:
- MVTec AD: The standard benchmark for industrial AD, containing 15 categories of objects and textures.
- VisA: A dataset consisting of 12 industrial object categories.
- MVTec LOCO AD: A specialized dataset designed to test "logical" anomalies (e.g., incorrect counts or arrangements) alongside standard structural ones.
Key Insights and Ideas
- Decision Logits > Text Generation: By using typed decision models, the authors avoid the latency and "hallucination" issues of autoregressive text generation. They simply extract the logits for "Yes/No" answers, making the process efficient and direct.
- The Power of Context: The "Box Decision" in LVR is a clever way to provide a VLM with spatial context without needing complex architectural changes. Seeing the whole image helps the model understand if a specific part is missing or misplaced.
- Refinement, Not Generation: LVR is designed to reweight existing anomaly maps. It can suppress noise, but it won't "invent" new anomaly evidence where the base detector found none. This makes it a safe, conservative enhancement for existing pipelines.
- Plug-and-Play Generality: The framework consistently improved performance regardless of whether the base detector was PatchCore, AnomalyDINO, or SubspaceAD.
Key Results
VerdictAD achieved state-of-the-art results across multiple settings. Notably, it provided a massive boost to image-level detection:
- MVTec LOCO AD: When applied to the SubspaceAD detector, VerdictAD improved the Image AUROC from 67.5% to 75.3% in a 4-shot setting.
- VisA: On the VisA dataset, it improved the Image AUROC of AnomalyDINO by 7.8 percentage points in the 1-shot setting.
- Qualitative Cleanup: As shown in the figure below, VerdictAD effectively suppresses "distracting" activations in the background or on normal parts of the object, resulting in much cleaner and more accurate localization maps.
The authors also demonstrated that VerdictAD helps resolve cases where normal variations (like juice color) trigger false alarms in standard detectors.
Limitations and Future Work
The primary limitation is that VerdictAD is a refiner. If the underlying base detector fails to find any evidence of a defect (i.e., the anomaly map is completely blank in the defective region), VerdictAD cannot recover that information. It relies on the base detector to provide at least some initial signal to work with.
Future work could involve exploring even more specialized decision models or finding ways to integrate the decision-making process more deeply into the feature extraction stage of the base detectors.
Summary updated