Keskeiset käsitteet
Maintaining re-identification performance in changing test environments through Test-time Similarity Modification.
Tiivistelmä
The content discusses the challenges of distribution shift in person re-identification and introduces TEMP, a novel Test-time Adaptation method. It addresses the limitations of existing methods by proposing re-id entropy as an uncertainty measure and demonstrates improved performance in online settings with changing distributions.
Structure:
Introduction to Person Re-identification Challenges
Existing Methods for Adaptation (UDA, TTA)
Proposal of TEMP Methodology
Experimental Results on Market-1501, MSMT17, and PersonX datasets
Comparison with Baseline Methods (No-adapt, BN-adapt, SourceTent, BNTA)
Sensitivity Analysis of Hyperparameter k
Visualization of Feature Space Alignment by TEMP
Tilastot
"TEMP improves the performance of re-id by up to about nine points compared with the TTA baselines under temporal distribution shifts in a batched online manner without accessing source data."
"TEMP aligns the clusters of the query and gallery features."
Lainaukset
"TEMP is the first fully TTA method for re-id that enables to reuse and adapt arbitrary off-the-shelf models trained in arbitrary ways during testing."
"We propose TEMP, a novel FTTA method for re-id."