ReadNet: Towards Accurate ReID with Limited and Noisy Samples

Yitian Li, Ruini Xue, Mengmeng Zhu, Qing Xu, Zenglin Xu

Abstract

Person re-identification (ReID) is a challenging cross-camera retrieval task to identify pedestrians. Many complex network structures are proposed recently and many of them concentrate on multi-branch features to achieve high performance. However, they are too heavy-weight to deploy in real-world applications. Additionally, pedestrian images are often captured by different surveillance cameras, so the varied lights, perspectives and resolutions result in inevitable multi-camera domain gaps for ReID. To address these issues, this paper proposes ATCN, a simple but effective angular triplet loss-based camera network, which is able to achieve compelling performance with only global features. In ATCN, a novel angular distance is introduced to learn a more discriminative feature representation in the embedding space. Meanwhile, a lightweight camera network is designed to transfer global features to more discriminative features. ATCN is designed to be simple and flexible so it can be easily deployed in practice. The experiment results on various benchmark datasets show that ATCN outperforms many SOTA approaches.

Introduction

Person re-identification (ReID)—the task of identifying the same individual across disjoint surveillance cameras—is a cornerstone of modern security and activity analysis. However, the task is notoriously difficult due to the "Camera Gap." A person's appearance can vary significantly when viewed from different angles, resolutions, or lighting conditions across a camera network.

Figure 1: Visual examples of the camera gap, showing how the same identity (e.g., P595) appears differently across cameras C6 and C3.

While many state-of-the-art (SOTA) models rely on complex, multi-branch architectures to extract local features, these are often too computationally expensive for real-world deployment. This paper introduces the Angular Triplet Camera Network (ATCN), a leaner approach that achieves high performance using only global features by optimizing the distance metric and employing adversarial feature transfer.

Methodology

The ATCN framework operates in two distinct stages to address the metric and camera gaps respectively.

1. The Feature-Extract Stage: Angular Triplet Loss

Standard ReID models often use Triplet Loss with Euclidean distance, which can be unstable. While Cosine Similarity is an alternative, its gradients diminish as the model improves, hindering fine-tuning. ATCN introduces Angular Triplet Loss ($L_{ang}$). By optimizing the angle between feature vectors rather than their straight-line distance, the model creates a more robust linear decision boundary. Crucially, the angular distance provides a steeper gradient than cosine similarity, allowing the model to continue learning effectively even as samples become well-separated.

2. The Feature-Transfer Stage: Adversarial Camera Network

To mitigate camera-specific noise, ATCN uses a lightweight adversarial setup. An Adapter network modifies global features to "trick" a Discriminator that is trained to identify the source camera. This minimax game forces the Adapter to strip away camera-specific information (like lighting tints or perspective bias) while retaining identity-critical features. This approach is significantly more efficient than image-to-image translation methods using GANs.

Figure 2: The ATCN architecture overview, illustrating the two-stage process and the mathematical advantage of angular distance over cosine similarity.

Experimental Results

ATCN was evaluated on the Market1501 and DukeMTMC-reID benchmarks, demonstrating that global features can rival complex part-based models when properly optimized.

  • Market1501: The model achieved 86.9% mAP and 94.5% Rank-1 accuracy, surpassing several multi-branch architectures.
  • DukeMTMC-reID: ATCN reached 77.1% mAP and 87.9% Rank-1 accuracy.
  • Efficiency: Compared to the popular Part-based Convolutional Baseline (PCB), ATCN requires approximately 15% less memory and offers 30-50% faster training times, making it highly suitable for practical applications.

Conclusion

ATCN demonstrates that accurate person re-identification can be achieved through a simple yet mathematically grounded approach. By replacing standard distance metrics with angular distance and utilizing a lightweight adversarial camera network, the model bridges the domain gap between cameras without the overhead of complex multi-branch networks. Future research will focus on enhancing the universality of these camera-invariant features for deployment in unseen environments.

Summary updated

Abstract

Person re-identification (ReID) is a challenging crosscamera retrieval task to identify pedestrians. Many complex network structures are proposed recently and many of them concentrate on multi-branch features to achieve high performance. However, they are too heavy-weight to deploy in realworld applications. Additionally, pedestrian images are often captured by different surveillance cameras, so the varied lights, perspectives and resolutions result in inevitable multi-camera domain gaps for ReID. To address these issues, this paper proposes ATCN, a simple but effective angular triplet loss-based camera network, which is able to achieve compelling performance with only global features. In ATCN, a novel angular distance is introduced to learn a more discriminative feature representation in the embedding space. Meanwhile, a lightweight camera network is designed to transfer global features to more discriminative features. ATCN is designed to be simple and flexible so it can be easily deployed in practice. The experiment results on various benchmark datasets show that ATCN outperforms many SOTA approaches.
BibTeX
@inproceedings{li2021readnet,
  author = {Yitian Li and Ruini Xue and Mengmeng Zhu and Qing Xu and Zenglin Xu},
  title = {{ReadNet: Towards Accurate ReID with Limited and Noisy Samples}},
  booktitle = {IJCNN},
  year = {2021},
  url = {https://openreview.net/forum?id=L99jCDL1ET},
}
Figure