Research map / Vision-language and multimodal / CLIP and zero-shot
CableInspect-AD: An Expert-Annotated Anomaly Detection Dataset
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces CableInspect-AD, an expert-annotated power line cable anomaly detection dataset from Hydro-Québec, and an Enhanced-PatchCore variant that sets thresholds from training nominal images, benchmarked against open VLMs in zero-shot settings.
Problem
Visual anomaly detection for robotic power line cable inspection lacks public, realistic datasets with diverse, graded anomalies, and existing methods need validation anomalies to set thresholds, which is impractical in deployment.
Approach
Enhanced-PatchCore computes the nearest-neighbour anomaly score of each training patch against the rest of the memory bank, so a detection threshold can be set from nominal training images alone; the paper also evaluates open conversational VLMs zero-shot with a yes/no prompt and VQAScore-based anomaly scores.
Key results
- CogVLM2-19B zero-shot reaches AUROC 0.86 and AUPR 0.91 with the lowest FPR (0.04), while CogVLM-17B has the best zero-shot F1 (0.77).
- Enhanced-PatchCore at 100 shots with beta-prime-95 thresholding reaches F1 0.75 and AUROC 0.78, competitive with the VLMs.
- Enhanced-PatchCore achieves AUPRO 0.53 on the cropped dataset versus 0.27 for WinCLIP; all models miss light-grade and subtle anomalies.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CogVLM-17B instead of the other evaluated zero-shot models | Zero-shot VAD · CableInspect-AD image-level | F1 score: 0.69 $\pm$ 0.02 -> 0.77 $\pm$ 0.02 | Helped |
| Use CogVLM2-19B instead of the other evaluated zero-shot models | Zero-shot VAD · CableInspect-AD image-level | AUPR: 0.84 $\pm$ 0.06 -> 0.91 $\pm$ 0.02 | Helped |
| Use CogVLM2-19B instead of CogVLM-17B | Zero-shot VAD · CableInspect-AD image-level | F1 score: 0.77 $\pm$ 0.02 -> 0.66 $\pm$ 0.04 | Hurt |
| Use Enhanced-PatchCore instead of CogVLM-17B | CableInspect-AD image-level detection · CableInspect-AD image-level | F1 score: 0.77 $\pm$ 0.02 -> 0.75 $\pm$ 0.03 | Hurt |
| Increase the number of Enhanced-PatchCore training images (ablation) | Enhanced-PatchCore · CableInspect-AD image-level | F1 score | Helped |
| Increase the number of Enhanced-PatchCore training images (ablation) | Enhanced-PatchCore · CableInspect-AD image-level | AUROC | Helped |
| Use max thresholding instead of the other tested threshold strategies (ablation) | Enhanced-PatchCore · CableInspect-AD image-level | F1 score | Hurt |
| Remove the image background by cropping to the central cable region (ablation) | Enhanced-PatchCore · CableInspect-AD cropped image-level | F1 score | Helped |
| Remove the image background by cropping to the central cable region (ablation) | Conversational VLMs · CableInspect-AD cropped image-level | F1 score | Hurt |
| Remove the image background by cropping to the central cable region (ablation) | WinCLIP · CableInspect-AD cropped image-level | AUROC | Helped |
| Use Enhanced-PatchCore instead of WinCLIP for pixel-level segmentation | CableInspect-AD cropped segmentation · CableInspect-AD cropped segmentation | AUPRO: $0.27$ $\pm$ $0.06$ -> $0.53$ $\pm$ $0.08$ | Helped |
Surprising findings
- Cropping out the background markedly reduced the performance of the conversational VLMs, despite improving Enhanced-PatchCore's stability.
- CogVLM2-19B has the highest AUROC and AUPR but a weaker F1, which the authors attribute to suboptimal thresholding.
Limitations
- Anomaly ratio in the dataset is higher than in real-world inspection, where anomalies are rare.
- The dataset does not cover every possible real-world anomaly, such as snow or bird droppings.
- Only three cables are included and results vary strongly across folds and cables.
Experimental setup
Datasets: CableInspect-AD MVTec AD
Benchmarks: CableInspect-AD CableInspect-AD_cropped
Metrics: F1 score FPR AUPR AUROC AUPRO precision recall FNR
Baselines: WinCLIP LLaVA 1.5-7B LLaVA 1.5-13B BakLLaVA-7B CogVLM-17B CogVLM2-19B PatchCore
Builds on: PatchCore WinCLIP VQAScore AnomalyGPT
Closest papers by meaning
- Supervised Anomaly Detection for Complex Industrial Images (CVPR 2024 · similarity 0.87)
- AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models (AAAI 2024 · similarity 0.86)
- Advancing Video Anomaly Detection: A Concise Review and a New Dataset (NeurIPS 2024 · similarity 0.85)
- ADPretrain: Advancing Industrial Anomaly Detection via Anomaly Representation Pretraining (NeurIPS 2025 · similarity 0.85)
- 3CAD: A Large-Scale Real-World 3C Product Dataset for Unsupervised Anomaly Detection (AAAI 2025 · similarity 0.84)
- Multi-Sensor Object Anomaly Detection: Unifying Appearance, Geometry, and Internal Properties (CVPR 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.