RGB Ensemble Strategies for Unsupervised Industrial Anomaly Detection on the AutoVI Dataset

Real automotive inspection lines need robust defect detection under cluttered backgrounds, fluctuating illumination, and operator-introduced clutter, conditions under which fully supervised pipelines are rarely feasible because defective samples are scarce and heterogeneous. We address this gap with a deployment-oriented study of unsupervised anomaly detection (UAD) on AutoVI, a public automotive benchmark, and we go beyond running existing detectors in three ways. First, we establish unified RGB and pseudo-depth baselines for seven UAD models under a single calibration and evaluation policy that combines threshold-agnostic metrics (AUROC, AP), operational metrics (TPR at fixed TNR), and pixel-level sPRO/AUsPRO under a 5% false-positive budget. Second, we show that a plug-and-play late fusion of independently calibrated RGB detectors consistently recovers pixel-level localization that no single model achieves, with no extra training and no architectural change; effects are large (Cohen’s d>2 on every flagged improvement) and statistically significant across three seeds. Third, we report an actionable negative result: combining RGB with monocular pseudo-depth through the same fusion scheme degrades rather than improves localization, and we trace the failure to the relative nature of estimated depth interacting with a parameter-free aggregator. Ablations on the fusion operator, ensemble size, and calibration support these findings, and the released calibrated artifacts make the comparison reproducible on other MVTec-style benchmarks.