Deep Nearest Neighbor Anomaly Detection

TL;DR

This paper compares deep self-supervised features with Imagenet pre-trained features for nearest neighbor anomaly detection, showing the latter's superior performance in accuracy, generalization, training time, and noise robustness.

Contribution

The study demonstrates that nearest neighbor methods on Imagenet pre-trained features outperform recent self-supervised deep methods for anomaly detection.

Findings

Nearest neighbor with Imagenet features outperforms self-supervised methods.

Pre-trained features lead to better accuracy and noise robustness.

The approach requires less training time and fewer assumptions.

Abstract

Nearest neighbors is a successful and long-standing technique for anomaly detection. Significant progress has been recently achieved by self-supervised deep methods (e.g. RotNet). Self-supervised features however typically under-perform Imagenet pre-trained features. In this work, we investigate whether the recent progress can indeed outperform nearest-neighbor methods operating on an Imagenet pretrained feature space. The simple nearest-neighbor based-approach is experimentally shown to outperform self-supervised methods in: accuracy, few shot generalization, training time and noise robustness while making fewer assumptions on image distributions.