Analysis of Pretrained Feature Extractors for Anomaly Detection

This technical report presents a systematic analysis of the influence of pretrained feature extractors on unsupervised anomaly detection (AD) methods. The study investigates various feature extractors, their intermediate layers, and pretraining protocols to understand their impact on anomaly detection performance. It highlights the sensitivity of existing AD methods to the choice of feature space, demonstrating that selecting an optimal feature layer can significantly enhance detection accuracy. The report outlines the importance of feature extractors in differentiating between normal and anomalous data, emphasizing that many current approaches lack a standardized feature selection strategy. The findings suggest that careful selection of feature layers is crucial for developing effective AD systems. Additionally, the report motivates further research into methods that can identify the best-performing feature layers based on specific datasets. This analysis serves as a foundation for advancing the understanding of feature extractor selection in the context of anomaly detection.