COSTARR: Consolidated Open Set Technique with Attenuation for Robust Recognition

Ryan Rabinowitz, Steve Cruz, Walter Scheirer, Terrance E. Boult · 2025

Read COSTARR: Consolidated Open Set Technique with Attenuation for Robust Recognition by Ryan Rabinowitz, Steve Cruz, Walter Scheirer, Terrance E. Boult (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-2508-01087.txt.

Description

Ryan Rabinowitz, Steve Cruz, Walter Scheirer, and Terrance Boult teach vision systems to spot what they have never seen, in a 2025 paper. Open-set recognition usually flags novelty by missing familiar features. The team proposes the opposite signal: training shrinks some weights, and those attenuated features carry the very information that separates known from unknown. Their method, COSTARR, demands familiar features and the absence of unfamiliar ones at once.

A probability reading ties the score to being both correct and known. Ablations show both raw and attenuated feature sets are essential. On large benchmarks with a thousand known classes, across several modern architectures, COSTARR beat prior leaders by clear margins. The authors say the field had been discarding half the evidence.

Catalog record

CREATORRyan Rabinowitz, Steve Cruz, Walter Scheirer, Terrance E. Boult
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE61 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-01087

Browse the full library · About the Tower of Babel