Research map / Architectures and efficiency / Attention and transformers
MambaAD: Exploring State Space Models for Multi-class Unsupervised Anomaly Detection
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.
Applies Mamba-based state space decoder (MambaAD: pretrained encoder plus LSS modules with HSS multi-directional scanning) to multi-class unsupervised anomaly detection, reporting SoTA on six datasets.
Problem
Multi-class unsupervised anomaly detection suffers from CNNs' weak long-range modeling and transformers' quadratic cost, limiting reconstruction-based methods like UniAD to the smallest feature maps.
Approach
A pyramid auto-encoder with a pretrained encoder and a Mamba decoder whose Locality-Enhanced State Space (LSS) modules combine cascaded Hybrid State Space (HSS) blocks (five scanning methods, eight directions, including Hilbert scan) with parallel multi-kernel convolutions, computing multi-scale anomaly maps.
Key results
- On MVTec-AD multi-class anomaly detection, MambaAD reports mean 98.6 AU-ROC / 99.6 AP / 97.8 F1_max, the best among compared methods.
- On VisA multi-class anomaly detection, MambaAD reports mean 94.3 AU-ROC / 94.5 AP, outperforming UniAD and DiAD in the appendix tables.
- On MVTec-AD localization, MambaAD reports mean 97.7 AU-ROC and 93.1 AU-PRO.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MambaAD instead of DiAD | multi-class anomaly detection · MVTec-AD image-level | AU-ROC: 97.2 -> 98.6 | Helped |
| Use MambaAD instead of DiAD | multi-class anomaly detection · MVTec-AD image-level | AP: 99.0 -> 99.6 | Helped |
| Use MambaAD instead of DiAD | multi-class anomaly detection · MVTec-AD image-level | F1_max: 96.5 -> 97.8 | Helped |
| Use MambaAD instead of SimpleNet | multi-class anomaly localization · MVTec-AD pixel-level | AU-ROC: 96.9 -> 97.7 | Helped |
| Use MambaAD instead of DeSTSeg | multi-class anomaly localization · MVTec-AD pixel-level | AP: 54.3 -> 56.3 | Helped |
| Use MambaAD instead of DiAD | multi-class anomaly localization · MVTec-AD pixel-level | F1_max: 55.5 -> 59.2 | Helped |
| Use MambaAD instead of RD4AD | multi-class anomaly localization · MVTec-AD pixel-level | AU-PRO: 91.1 -> 93.1 | Helped |
| Use MambaAD instead of RD4AD | multi-class anomaly detection · VisA image-level | AU-ROC: 92.4 -> 94.3 | Helped |
| Use MambaAD instead of RD4AD | multi-class anomaly detection · VisA image-level | AP: 92.4 -> 94.5 | Helped |
| Use MambaAD instead of RD4AD | multi-class anomaly detection · VisA image-level | F1_max: 89.6 -> 89.4 | Hurt |
| Use MambaAD instead of RD4AD | multi-class anomaly localization · VisA pixel-level | AU-ROC: 98.1 -> 98.5 | Helped |
| Use MambaAD instead of DeSTSeg | multi-class anomaly localization · VisA pixel-level | AP: 39.6 -> 39.4 | Hurt |
Limitations
- Not stated in the provided text
Experimental setup
Datasets: MVTec-AD VisA MVTec-3D Uni-Medical COCO-AD Real-IAD
Benchmarks: MVTec-AD VisA MVTec-3D Uni-Medical COCO-AD Real-IAD
Metrics: AU-ROC AP F1_max AU-PRO
Baselines: RD4AD UniAD SimpleNet DeSTSeg DiAD ViTAD
Builds on: Mamba VMamba RD4AD UniAD Hilbert curve scanning
Closest papers by meaning
- ResAD: A Simple Framework for Class Generalizable Anomaly Detection (NeurIPS 2024 · similarity 0.86)
- Long-Tailed Anomaly Detection with Learnable Class Names (CVPR 2024 · similarity 0.86)
- Hyperbolic Anomaly Detection (CVPR 2024 · similarity 0.86)
- Dinomaly: The Less Is More Philosophy in Multi-Class Unsupervised Anomaly Detection (CVPR 2025 · similarity 0.85)
- Swift Hydra: Self-Reinforcing Generative Framework for Anomaly Detection with Multiple Mamba Models (ICLR 2025 · similarity 0.84)
- AnomalyNCD: Towards Novel Anomaly Class Discovery in Industrial Scenarios (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.