Uncertainty-Aware Likelihood Ratio Estimation for Pixel-Wise Out-of-Distribution Detection

Marc Hölle, Walter Kellermann, Vasileios Belagiannis · 2025

Read Uncertainty-Aware Likelihood Ratio Estimation for Pixel-Wise Out-of-Distribution Detection by Marc Hölle, Walter Kellermann, Vasileios Belagiannis (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-00587.txt.

Description

Self-driving cars must spot the odd thing they were never taught, without crying wolf at rare known ones. Marc Hölle, Walter Kellermann, and Vasileios Belagiannis tackle that pixel-by-pixel task in 2025. Their method weighs the odds that a pixel is known against the odds it is strange, and it keeps track of its own doubt while doing so. Doubt comes from thin training cases and from fake odd cases that stand in for real ones.

Rather than one hard score, the method hands out a spread of scores that shows how sure it is. On five known test sets it posted the lowest false-alarm rate of its class, about two and a half in a hundred. Its hit rate stayed high at the same time, and the extra math cost next to nothing. The code is public.

Catalog record

CREATORMarc Hölle, Walter Kellermann, Vasileios Belagiannis
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE40 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-00587

Browse the full library · About the Tower of Babel