Belangrijkste concepten
The author proposes a Generalist Anomaly Detection (GAD) model, InCTRL, that utilizes few-shot normal images as sample prompts to detect anomalies across diverse datasets without further training.
Samenvatting
The paper introduces InCTRL, a GAD model trained on auxiliary data to detect anomalies in various domains using few-shot normal image prompts. It outperforms competing methods on industrial defects, medical anomalies, and semantic anomalies in both one-vs-all and multi-class settings.
The approach involves learning an in-context residual model for GAD named InCTRL, which discriminates anomalies from normal samples based on residuals between query images and few-shot normal sample prompts. This enables the model to generalize across different datasets without additional training.
Experiments on nine diverse AD datasets establish a GAD benchmark covering industrial defect detection, medical image anomaly detection, and semantic anomaly detection. InCTRL significantly outperforms state-of-the-art methods across all datasets.
InCTRL's success is attributed to its ability to capture fine-grained residuals at both patch and image levels while incorporating text prompt-guided prior knowledge. The model demonstrates remarkable generalization capabilities across various anomaly detection tasks.
Statistieken
False positive rate: 0.733±0.000
True positive rate: 0.946±0.000
Baseline AUROC: 0.855±0.000
Citaten
"InCTRL significantly outperforms state-of-the-art competing methods."
"Our results show that InCTRL achieves the best performance."