Reproduce “ShiftSplit-AD: Separating Domain Shift from Defects in Foundation-Feature Visual 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
Muhamathu Ameer Ali Aacaas Muhamath
ShiftSplit-AD introduces a post-hoc, training-free method that decomposes nearest-normal DINOv2 residuals into low-rank and row-sparse components to differentiate domain shifts from defects in visual anomaly detection. The approach improved image AUROC from 0.6780 to 0.7294 on the AeBAD-S real-world domain shift dataset but exhibited heterogeneous performance across defect types and reduced pixel-level localization accuracy on MVTec Bottle.
Key insights
- The low-rank and sparse components of the residual matrix do not hold universally distinct semantic meanings (e.g., low-rank is not exclusively domain shift, and sparse is not exclusively defect).
- The approach acts as a reweighting mechanism, and its benefits are conditional, depending on whether the filtered residual structure constitutes nuisance (domain shift) or crucial defect signal.
- There is an inherent trade-off: aggressive sparse filtering can suppress broad, correlated domain shift evidence but may inadvertently remove useful defect information, particularly for defects with structured patterns or when raw features are already discriminative.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.