Reproduce “SPARC: Subspace Position-Aware Robust Few-Shot Calibration for Distribution-Shifted Industrial 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
Seokhee Han, Seungjun Chu, Mateusz Nowak, Peter Chin
SPARC introduces a gradient-free, few-shot calibration method for industrial anomaly detection, addressing performance degradation from deployment shifts by employing per-cell subspace projection to remove nuisance. The method improved Image AUROC by an average of 13.8 percentage points and AU-PRO by 3.5 percentage points on shift-prone benchmarks, utilizing only k 8 normal images.
Key insights
- Distribution shifts in industrial anomaly detection can be effectively modeled and removed as spatially varying, low-dimensional nuisance components within per-cell feature subspaces.
- Calibrating feature representations through per-cell subspace projection is more effective at removing deployment-time nuisance than global statistical alignment or simple mean-only corrections.
- Robust, few-shot calibration can be achieved efficiently without gradient-based updates or extensive hyperparameter tuning by leveraging the algebraic properties of feature disagreement among a small set of normal samples.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.