# A noisy elephant in the room: Is your out-of-distribution detector robust to label noise?

**Authors:** Galadrielle Humblot-Renaux, Sergio Escalera, Thomas B. Moeslund  
**arXiv:** [2404.01775](https://arxiv.org/abs/2404.01775)  
**Date:** 2024-04-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2404.01775v1/

## Key points

- Computer vision requires detecting unfamiliar or unexpected images for safe automated system deployment
- Out-of-distribution (OOD) detection is crucial for identifying images outside a model's training domain
- Limited discussion on OOD detection performance with unreliable labels in the training dataset
- Study by Galadrielle Humblot-Renaux, Sergio Escalera, and Thomas B. Moeslund investigates 20 OOD detection methods under label noise conditions
- Poor separation between incorrectly classified in-distribution (ID) samples and OOD samples is a key limitation of existing methods with label noise
- Importance of robustness in OOD detection systems and considering label quality during model training

## Summary

, , , , 

In the field of computer vision, the ability to detect unfamiliar or unexpected images is crucial for ensuring the safe deployment of automated systems. One key aspect of this is out-of-distribution (OOD) detection, which involves identifying images that fall outside of a model's training domain. While there has been significant research interest in developing methods for OOD detection, there has been limited discussion on how these methods perform when the underlying classifier is trained on a dataset with unreliable labels. In their work titled "A noisy elephant in the room: Is your out-of-distribution detector robust to label noise? ", authors Galadrielle Humblot-Renaux, Sergio Escalera, and Thomas B. Moeslund address this gap by investigating 20 state-of-the-art OOD detection methods in a more realistic scenario where the labels used to train the classifier are potentially noisy (e.g., crowd-sourced or web-scraped). Through extensive experiments involving different datasets, levels and types of noise, various architectures, and checkpointing strategies, they provide valuable insights into how class label noise impacts OOD detection performance. The study reveals that poor separation between incorrectly classified in-distribution (ID) samples and OOD samples is an important limitation of existing methods when dealing with label noise. By shedding light on this overlooked challenge, the authors contribute to a deeper understanding of OOD detection under real-world conditions. Their findings not only highlight the need for robustness in OOD detection systems but also underscore the importance of considering label quality during model training. The code for their research is available at https://github.com/glhr/ood-labelnoise. This work was accepted at CVPR 2024 and falls under primary categories cs.CV, cs.AI, and cs.LG.