Reproduce “EfficientAD: Accurate Visual Anomaly Detection at Millisecond-Level Latencies” 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
Kilian Batzner, Lars Heckler, Rebecca König
MVTec Software GmbH researchers developed EfficientAD, a visual anomaly detection system that achieves state-of-the-art detection and localization accuracy while operating at millisecond-level latencies. The system, available in 'S' and 'M' variants, demonstrates superior performance on MVTec AD, VisA, and MVTec LOCO datasets, processing images at up to 614 frames per second while maintaining high detection rates for both structural and logical anomalies.
Key insights
- Loss-induced asymmetry, achieved through novel loss functions like hard feature loss and pretraining penalty, can significantly boost student-teacher model performance without increasing inference runtime.
- A highly efficient and compact Patch Description Network (PDN), distilled from a large pre-trained model, can extract expressive features at millisecond-level latencies while maintaining good localization due to its controlled receptive field.
- Combining a local anomaly detection mechanism (student-teacher on PDN features) with a global autoencoder-based mechanism, efficiently integrated via a shared student network, enables robust detection of both structural and logical anomalies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.