DisCoPatch: Batch Statistics Are All You Need For OOD Detection, But Only If You Can Trust Them

Francisco Caetano, Christiaan Viviers, Luis A. Zavala-Mondragón, Peter H.N. de With, Fons van der Sommen · 2025

Read DisCoPatch: Batch Statistics Are All You Need For OOD Detection, But Only If You Can Trust Them by Francisco Caetano, Christiaan Viviers, Luis A. Zavala-Mondragón, Peter H.N. de With, Fons van der Sommen (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-2501-08005.txt.

Description

Francisco Caetano and co-authors present DisCoPatch, a method to spot data that falls outside a model's training range in a 2025 paper. Their focus is on subtle shifts in data traits, which can quietly degrade model performance. In trained discriminator models, real and fake samples form distinct groups with distinct batch statistics, a trait the team exploits.

The method uses an unsupervised model that rebuilds images, and its flawed outputs serve as fake samples to train the discriminator. At test time, batches are patches from one image, so batch statistics stay consistent and trustworthy. The tighter boundary between normal and shifted data gave top scores on public test sets.

Catalog record

CREATORFrancisco Caetano, Christiaan Viviers, Luis A. Zavala-Mondragón, Peter H.N. de With, Fons van der Sommen
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-08005

Browse the full library · About the Tower of Babel