Research map / Vision-language and multimodal / CLIP and zero-shot
MetaUAS: Universal Anomaly Segmentation with One-Prompt Meta-Learning
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.
Recasts zero/one-shot visual anomaly segmentation as change segmentation and trains a one-prompt meta-learning model (MetaUAS) on synthetic COCO-based image pairs, with a soft feature alignment module, achieving state-of-the-art results on MVTec, VisA and Goods without language guidance.
Problem
Universal visual anomaly segmentation for unseen objects and anomalies, using only one normal image prompt and without target anomaly datasets or vision-language models.
Approach
Reformulate anomaly segmentation as change segmentation between a normal prompt and a query, train on large-scale synthetic image pairs with object-level (inpainting, pasting) and local (Perlin-noise DRAEM-style) changes, and align prompt and query features with a soft cross-attention alignment module before a UNet decoder.
Key results
- On MVTec, MetaUAS with one random normal prompt reaches P-PR 59.3 and P-PRO 82.6, and MetaUAS*+ reaches P-ROC 97.6 and P-PRO 92.5, above WinCLIP+ (P-PR 38.4).
- On VisA, MetaUAS*+ achieves P-ROC 98.0 and P-PRO 85.5 versus AnomalyCLIP's P-ROC 95.5 and P-PRO 86.8 (red-marked best on PRO).
- MetaUAS runs in 3.1 ms per image at 256x256 with 22.1M total parameters (4.6M learnable), versus roughly 340 ms for WinCLIP+ with 10x more parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add soft feature alignment rather than no alignment (ablation) | MetaUAS · MVTec | P-PR: 44.9 -> 59.6 | Helped |
| add soft instead of hard feature alignment (ablation) | MetaUAS · MVTec | P-PR: 48.2 -> 59.6 | Helped |
| use feature concatenation instead of element-wise addition (ablation) | MetaUAS with soft alignment · MVTec | P-PR: 24.0 -> 59.6 | Helped |
| use feature concatenation instead of absolute difference (ablation) | MetaUAS with soft alignment · MVTec | P-PR: 45.9 -> 59.6 | Helped |
| freeze rather than train the encoder (ablation) | MetaUAS with EfficientNet-b4 · MVTec | P-PR: 50.3 -> 59.6 | Helped |
| replace EfficientNet-b4 with EfficientNet-b6 (ablation) | MetaUAS · MVTec | P-PR: 59.6 -> 56.9 | Hurt |
| replace EfficientNet-b4 with EfficientViT-b3 (ablation) | MetaUAS · MVTec | P-PR: 59.6 -> 58.5 | Hurt |
| replace EfficientNet-b4 with MobileNetV2 (ablation) | MetaUAS · MVTec | P-PR: 59.6 -> 33.7 | Hurt |
| include object-level and local-region changes rather than only local-region changes (ablation) | MetaUAS training data · MVTec | P-PR: 44.3 -> 59.6 | Helped |
| include local-region changes rather than only object-level changes (ablation) | MetaUAS training data · MVTec | P-PRO: 75.4 -> 82.6 | Helped |
| replace UNet with FPN-Cat (ablation) | MetaUAS decoder · MVTec | P-PR: 59.6 -> 49.9 | Hurt |
| replace UNet with FPN-Add (ablation) | MetaUAS decoder · MVTec | P-PR: 59.6 -> 51.4 | Hurt |
Surprising findings
- Freezing the EfficientNet-b4 encoder works better than end-to-end training, which the authors attribute to overfitting the change-segmentation data.
- Object-level synthetic changes contribute more than local-region changes alone, and Add fusion of query and prompt features performs worst.
Limitations
- Performance depends on choosing an appropriate normal prompt; cosine-similarity prompt retrieval is used when class is unknown.
- Struggles on fine-grained categories where a classifier may be needed to pick the prompt.
- Weaker on Goods, where one prompt per group does not match many subcategories.
Experimental setup
Datasets: MS-COCO MVTec VisA Goods
Benchmarks: MVTec VisA Goods
Metrics: I-ROC I-PR I-F1max P-ROC P-PR P-F1max P-PRO inference time parameter count
Baselines: CLIP WinCLIP WinCLIP+ AnomalyCLIP PatchCore UniAD
Builds on: UniAD DRAEM CYWS PatchCore WinCLIP
Closest papers by meaning
- A Dataset for Semantic Segmentation in the Presence of Unknowns (CVPR 2025 · similarity 0.88)
- Adapting Visual-Language Models for Generalizable Anomaly Detection in Medical Images (CVPR 2024 · similarity 0.87)
- Promptable Anomaly Segmentation with SAM Through Self-Perception Tuning (AAAI 2025 · similarity 0.87)
- UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.87)
- Unseen Visual Anomaly Generation (CVPR 2025 · similarity 0.87)
- One-to-Normal: Anomaly Personalization for Few-shot Anomaly Detection (NeurIPS 2024 · similarity 0.86)
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.