Better than Average: Spatially-Aware Aggregation of Segmentation Uncertainty Improves Downstream Performance

Vanessa Emanuela Guarino 1 , 2 , 5 , ∗ , 🖂 {}^{1,2,5,*,\text{\Letter}} Claudia Winklmayr 1,2,5,∗ Jannik Franzen 1,2,3,5,∗ Josef Lorenz Rumberger 1,2,3,4 Manuel Pfeuffer 4 Sonja Greven 4 Klaus Maier-Hein 2,6,7 Carsten T. Lüth 2 , 6 , 7 , † {}^{2,6,7,\text{\textdagger}} Christoph Karg 1 , 2 , † {}^{1,2,\text{\textdagger}} Dagmar Kainmueller 1 , 2 , 5 , † , 🖂 {}^{1,2,5,\text{\textdagger},\text{\Letter}} and 1 Max-Delbrück-Center (MDC) 2 Helmholtz Imaging 3 Charité Universitätsmedizin 4 Humboldt-Universität zu Berlin 5 University of Potsdam 6 German Cancer Research Center (DKFZ) 7 Heidelberg University and 🖂 {}^{\text{\Letter}} {firstnames.surname}@mdc-berlin.de ∗ / † {}^{*/\text{\textdagger}} equal contribution · 2026

Read Better than Average: Spatially-Aware Aggregation of Segmentation Uncertainty Improves Downstream Performance by Vanessa Emanuela Guarino 1 , 2 , 5 , ∗ , 🖂 {}^{1,2,5,*,\text{\Letter}} Claudia Winklmayr 1,2,5,∗ Jannik Franzen 1,2,3,5,∗ Josef Lorenz Rumberger 1,2,3,4 Manuel Pfeuffer 4 Sonja Greven 4 Klaus Maier-Hein 2,6,7 Carsten T. Lüth 2 , 6 , 7 , † {}^{2,6,7,\text{\textdagger}} Christoph Karg 1 , 2 , † {}^{1,2,\text{\textdagger}} Dagmar Kainmueller 1 , 2 , 5 , † , 🖂 {}^{1,2,5,\text{\textdagger},\text{\Letter}} and 1 Max-Delbrück-Center (MDC) 2 Helmholtz Imaging 3 Charité Universitätsmedizin 4 Humboldt-Universität zu Berlin 5 University of Potsdam 6 German Cancer Research Center (DKFZ) 7 Heidelberg University and 🖂 {}^{\text{\Letter}} {firstnames.surname}@mdc-berlin.de ∗ / † {}^{*/\text{\textdagger}} equal contribution (2026) 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-2603-29941.txt.

Description

Guarino and co-authors study how to roll pixel-level uncertainty into one score per image. Segmentation tools output a score per pixel, but downstream checks need image-level values. The global average is the default, yet it ignores spatial structure. The team analyzes flaws in common aggregation strategies. They propose new ones that use spatial uncertainty patterns. Benchmarks cover out-of-distribution and failure detection on ten datasets. Spatial methods win overall, though dataset traits matter. A meta-aggregator stays robust across sets.

Catalog record

CREATORVanessa Emanuela Guarino 1 , 2 , 5 , ∗ , 🖂 {}^{1,2,5,*,\text{\Letter}} Claudia Winklmayr 1,2,5,∗ Jannik Franzen 1,2,3,5,∗ Josef Lorenz Rumberger 1,2,3,4 Manuel Pfeuffer 4 Sonja Greven 4 Klaus Maier-Hein 2,6,7 Carsten T. Lüth 2 , 6 , 7 , † {}^{2,6,7,\text{\textdagger}} Christoph Karg 1 , 2 , † {}^{1,2,\text{\textdagger}} Dagmar Kainmueller 1 , 2 , 5 , † , 🖂 {}^{1,2,5,\text{\textdagger},\text{\Letter}} and 1 Max-Delbrück-Center (MDC) 2 Helmholtz Imaging 3 Charité Universitätsmedizin 4 Humboldt-Universität zu Berlin 5 University of Potsdam 6 German Cancer Research Center (DKFZ) 7 Heidelberg University and 🖂 {}^{\text{\Letter}} {firstnames.surname}@mdc-berlin.de ∗ / † {}^{*/\text{\textdagger}} equal contribution
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE113250
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.29941

Browse the full library · About the Tower of Babel