The paper presents a framework for evaluating image-based anomaly detectors using synthetically generated validation data. This method, which does not require additional training or fine-tuning, uses a small support set of normal images to create synthetic anomalies. When mixed with normal samples, these synthetic anomalies provide a validation framework for evaluating the accuracy of the anomaly detector and selecting the most effective models. The study found that this synthetic validation framework selects the same models and hyper-parameters as a ground-truth validation set. Furthermore, the study found that the use of the synthetic validation method in CLIP-based anomaly detection outperforms other strategies and leads to the best detection accuracy.
Publication date: 16 Oct 2023
Project Page: https://arxiv.org/abs/2310.10461
Paper: https://arxiv.org/pdf/2310.10461
