DynaMix: Generalizable Person Re-identification via Dynamic Relabeling and Mixed Data Sampling

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Mamedov, Timur, Konushin, Anton, Konushin, Vadim
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866914219729879040
author Mamedov, Timur
Konushin, Anton
Konushin, Vadim
author_facet Mamedov, Timur
Konushin, Anton
Konushin, Vadim
contents Generalizable person re-identification (Re-ID) aims to recognize individuals across unseen cameras and environments. While existing methods rely heavily on limited labeled multi-camera data, we propose DynaMix, a novel method that effectively combines manually labeled multi-camera and large-scale pseudo-labeled single-camera data. Unlike prior works, DynaMix dynamically adapts to the structure and noise of the training data through three core components: (1) a Relabeling Module that refines pseudo-labels of single-camera identities on-the-fly; (2) an Efficient Centroids Module that maintains robust identity representations under a large identity space; and (3) a Data Sampling Module that carefully composes mixed data mini-batches to balance learning complexity and intra-batch diversity. All components are specifically designed to operate efficiently at scale, enabling effective training on millions of images and hundreds of thousands of identities. Extensive experiments demonstrate that DynaMix consistently outperforms state-of-the-art methods in generalizable person Re-ID.
format Preprint
id arxiv_https___arxiv_org_abs_2511_19067
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle DynaMix: Generalizable Person Re-identification via Dynamic Relabeling and Mixed Data Sampling
Mamedov, Timur
Konushin, Anton
Konushin, Vadim
Computer Vision and Pattern Recognition
Artificial Intelligence
Machine Learning
Generalizable person re-identification (Re-ID) aims to recognize individuals across unseen cameras and environments. While existing methods rely heavily on limited labeled multi-camera data, we propose DynaMix, a novel method that effectively combines manually labeled multi-camera and large-scale pseudo-labeled single-camera data. Unlike prior works, DynaMix dynamically adapts to the structure and noise of the training data through three core components: (1) a Relabeling Module that refines pseudo-labels of single-camera identities on-the-fly; (2) an Efficient Centroids Module that maintains robust identity representations under a large identity space; and (3) a Data Sampling Module that carefully composes mixed data mini-batches to balance learning complexity and intra-batch diversity. All components are specifically designed to operate efficiently at scale, enabling effective training on millions of images and hundreds of thousands of identities. Extensive experiments demonstrate that DynaMix consistently outperforms state-of-the-art methods in generalizable person Re-ID.
title DynaMix: Generalizable Person Re-identification via Dynamic Relabeling and Mixed Data Sampling
topic Computer Vision and Pattern Recognition
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2511.19067