Abstract
Person Re-identification (Re-ID) typically relies on models pre-trained on ImageNet, which suffers from a significant domain gap. While unsupervised pre-training on person-specific data has shown promise, it often ignores temporal cues in video. This paper introduces LUPerson-NL, a massive dataset of 10 million images with 430,000 identities automatically labeled via video tracklets. To address the inherent noise in these labels, we propose the Pre-training with Noisy Labels (PNL) framework. PNL jointly learns feature representations and rectifies labels through prototype-based and label-guided contrastive learning. Our approach significantly outperforms ImageNet-supervised and unsupervised pre-training across major benchmarks, particularly in low-data regimes.
Introduction
The performance of Person Re-identification (Re-ID) systems is heavily dependent on the quality and scale of training data. However, manually labeling millions of images is prohibitively expensive. Existing methods often leverage ImageNet pre-training, but the semantic gap between generic objects and human pedestrians limits its effectiveness. Recent efforts in unsupervised pre-training (e.g., LUPerson) mitigate this gap but fail to utilize the "free" temporal supervision available in raw video tracklets.
This work bridges this gap by utilizing automatically generated, albeit noisy, labels from video tracking systems. We present LUPerson-NL, a large-scale dataset derived from 21,000 video scenes, and the PNL framework designed to learn robust representations from this noisy data.
The LUPerson-NL Dataset
Using the FairMOT tracker on street-view videos, the authors extracted over 10 million images. Each tracklet is assigned an identity, but tracking errors introduce two primary types of noise:
- Noise-I (Fragmentation): A single individual is split into multiple IDs.
- Noise-II (Drift): Different individuals are incorrectly assigned the same ID.
Despite these errors, the sheer scale and domain relevance of LUPerson-NL provide a superior foundation for Re-ID compared to generic datasets.
The PNL Framework
The PNL framework addresses label noise through three integrated components:
- Supervised Re-ID Learning: A classification module that learns from current labels.
- Prototype-based Contrastive Learning: The model maintains identity centroids (prototypes). By comparing image features to these prototypes, the system identifies and rectifies label errors.
- Label-guided Contrastive Learning: An extension of contrastive learning that uses rectified labels to refine the feature space, pulling similar identities together while pushing dissimilar ones apart.
Experimental Results
We evaluated PNL-pre-trained models on standard benchmarks including Market1501, DukeMTMC-reID, MSMT17, and CUHK03.
Performance and Scaling
PNL consistently outperforms both ImageNet-supervised and unsupervised LUPerson pre-training. On CUHK03, PNL achieves a 5.7% mAP improvement over unsupervised methods. As shown in the scaling analysis, the performance gap is most pronounced when downstream training data is scarce. For instance, with only 10% of the Market1501 training set, PNL improves mAP by 15.6% over unsupervised pre-training.
Key Insights
- Self-Correction: The "rectification loop" allows the model to improve its own labels as its feature representation matures. As prototypes become more accurate, label rectification becomes more reliable.
- Domain Alignment: Pre-training on person-centric data, even with noise, is more effective than high-quality pre-training on unrelated domains like ImageNet.
Conclusion
LUPerson-NL and the PNL framework demonstrate that large-scale, noisy, automatically-labeled video data is a powerful resource for Re-ID. By actively rectifying labels during pre-training, PNL achieves state-of-the-art results and shows remarkable robustness in data-constrained scenarios. Future work will explore more sophisticated tracking algorithms to further improve the quality of initial labels.
Summary updated