Self-Supervised Partial Cycle-Consistency for Multi-View Matching

Fedor Taggenbrock 1,2 , Gertjan Burghouts 1 and Ronald Poppe 2 1 Utrecht University, Utrecht, Netherlands 2 TNO, The Hague, Netherlands { fedor.taggenbrock, gertjan.burghouts } @tno.nl, [email protected] · 2025

Read Self-Supervised Partial Cycle-Consistency for Multi-View Matching by Fedor Taggenbrock 1,2 , Gertjan Burghouts 1 and Ronald Poppe 2 1 Utrecht University, Utrecht, Netherlands 2 TNO, The Hague, Netherlands { fedor.taggenbrock, gertjan.burghouts } @tno.nl, [email protected] (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-2501-06000.txt.

Description

Fedor Taggenbrock, Gertjan Burghouts and Ronald Poppe extend cycle-consistency training to matching objects across partly overlapping camera views in 2025. Multi-camera systems need features that stay the same across views, and labeling training data by hand is laborious.

Their math now handles partial overlap directly, and a pseudo-mask steers the loss to count only the shared region. New cycle variants and a time-spread sampling scheme for scenes add further gains. On a hard tracking data set, their method beat the prior self-supervised best by a clear margin in F1 score. Gains held up when overlap in training data was reduced.

Catalog record

CREATORFedor Taggenbrock 1,2 , Gertjan Burghouts 1 and Ronald Poppe 2 1 Utrecht University, Utrecht, Netherlands 2 TNO, The Hague, Netherlands { fedor.taggenbrock, gertjan.burghouts } @tno.nl, [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE34 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2501-06000

Browse the full library · About the Tower of Babel