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

2025-11-24

AI Summary

To address the limited generalization capability of person re-identification (Re-ID) across unseen cameras and environments, this paper proposes an end-to-end training framework that jointly leverages a small amount of multi-camera labeled data and large-scale single-camera pseudo-labeled data. Methodologically, it introduces: (1) a dynamic pseudo-label refinement mechanism integrating confidence-adaptive selection and cross-camera consistency constraints; (2) an efficient centroid maintenance strategy enabling stable metric learning on million-scale images and hundred-thousand-identity datasets; and (3) a hierarchical hybrid sampling scheme balancing supervision from labeled data and diversity from pseudo-labels. Extensive experiments demonstrate significant improvements over state-of-the-art methods on multiple cross-domain benchmarks. Notably, this is the first work to achieve robust, scalable generalization training at the million-person scale—offering both practical deployability and computational efficiency.

Technology Category

• Computer Vision: Multi-modal Vision
• Machine Learning: Multi-class/Multi-label Learning & Extreme Classification
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Security and Privacy: Large-scale security measurements

Abstract

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.

Problem

• Recognizing individuals across unseen cameras and environments
• Overcoming reliance on limited labeled multi-camera training data
• Effectively combining manually labeled and pseudo-labeled single-camera data

Innovation

• Dynamic relabeling module refines pseudo-labels on-the-fly
• Efficient centroids module maintains robust identity representations
• Data sampling module balances learning complexity and diversity