Reproduce “MambaADv2: Evolving Duality-enhanced State Space Model for Unsupervised Anomaly Detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiaobin Hu, Haoyang He, Bo Yin, Yu He, Lei Xie, Jiangning Zhang, Yu-Gang Jiang, Shuicheng Yan
MambaADv2, an unsupervised anomaly detection framework, leverages evolved Duality-enhanced State Space Models to achieve state-of-the-art performance across six diverse datasets, including MVTec-AD (87.2 mAD), by precisely reconstructing normal features and amplifying anomalies with linear computational complexity.
Key insights
- Evolving State Space Models to Mamba-3, combined with Rotary Position Embedding and State Space Duality, significantly enhances spatial information preservation and robust context modeling with linear complexity for anomaly detection.
- Explicitly integrating both spatial and frequency domain analysis through Wavelet Convolutions and a specialized Inception Mixer allows for a more comprehensive capture of diverse anomaly types, including high-frequency textural and low-frequency structural defects.
- Adapting scanning strategies (SAPS) to the semantic hierarchy of features improves both the efficiency and accuracy of sequence modeling in visual anomaly detection by avoiding redundancy in deeper, more abstract feature layers.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.