Runtime Monitoring for Out-of-Distribution Detection in Object Detection Neural Networks

TL;DR

This paper extends runtime monitoring techniques to object detection neural networks for effective out-of-distribution input detection, demonstrating their practical utility in industrial perception systems.

Contribution

It adapts a runtime-monitoring approach from classification to object detection, enabling OOD detection in perception systems with multiple objects.

Findings

Effective OOD detection across various scenarios

Demonstrated applicability in real-world perception systems

Extended runtime monitoring to complex object detection tasks

Abstract

Runtime monitoring provides a more realistic and applicable alternative to verification in the setting of real neural networks used in industry. It is particularly useful for detecting out-of-distribution (OOD) inputs, for which the network was not trained and can yield erroneous results. We extend a runtime-monitoring approach previously proposed for classification networks to perception systems capable of identification and localization of multiple objects. Furthermore, we analyze its adequacy experimentally on different kinds of OOD settings, documenting the overall efficacy of our approach.