Angular Triplet Loss-based Camera Network for ReID

Yitian Li, Ruini Xue, Mengmeng Zhu, Jing 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) involves identifying the same individual across a network of non-overlapping surveillance cameras. This task is complicated by "multi-camera domain gaps"—variations in lighting, camera angles, and image resolutions that make the same person look different across cameras.

Pedestrian images across different cameras illustrating the domain gap (P-ID, C-Camera).

Current state-of-the-art models often rely on heavy, multi-branch architectures that extract local features from specific body parts. While effective, these models are computationally intensive and difficult to deploy in real-time systems. Furthermore, standard training objectives like Euclidean triplet loss often struggle with convergence and compatibility with classification losses.

Proposed Methodology: ATCN

The proposed Angular Triplet-based Camera Network (ATCN) addresses these challenges through two primary components: a novel loss function and a lightweight adversarial camera network.

Angular Triplet Loss (AT)

Traditional ReID models use Euclidean ($L_2$) distance for triplet loss. However, classification layers typically employ softmax, which learns linear decision boundaries in an angular space. ATCN introduces angular distance to align the triplet loss with the classification objective. As shown in the distance plots, the angular distance provides a "sharper" gradient, facilitating better feature separation even when samples are already relatively well-clustered.

Camera Network (CN)

To mitigate camera-specific noise, ATCN employs a lightweight adversarial sub-network. This consists of a Camera Discriminator that attempts to predict the source camera of a given feature, and a Feature Transfer Adapter that modifies global features to "fool" the discriminator. This adversarial training forces the feature extractor to produce camera-invariant representations, focusing on identity rather than environmental artifacts.

The two-stage ATCN framework: (a) feature extraction and (b) feature transfer, alongside activation and distance function comparisons.

Experimental Results

ATCN was evaluated on the Market1501 and DukeMTMC-reID datasets, demonstrating that a well-refined global feature can outperform many multi-branch heavyweights.

Performance and Efficiency

The model achieved competitive results using only global features:

  • Market1501: 86.9% mAP and 94.5% Rank-1 accuracy.
  • DukeMTMC-reID: 77.1% mAP and 87.9% Rank-1 accuracy.

Crucially, ATCN is significantly more efficient than part-based models. It required only 6.33 GB of memory and 109 minutes for training, whereas similar high-performing models required up to 15 GB of memory and over 200 minutes of training time.

Ablation Study

Ablation experiments confirmed that both the Angular Triplet (AT) loss and the Camera Network (CN) contribute independently to performance gains, with the best results achieved when both are integrated.

Conclusion

ATCN demonstrates that high-performance ReID is achievable using a lightweight, global-feature-based approach. By optimizing the geometry of the loss function and employing adversarial feature cleaning, ATCN bridges the multi-camera domain gap while remaining efficient enough for practical deployment. Future work will explore more complex adapter architectures and the application of angular losses to broader cross-domain retrieval tasks.

Summary updated

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 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{li2021angular,
  author = {Yitian Li and Ruini Xue and Mengmeng Zhu and Jing Xu and Zenglin Xu},
  title = {{Angular Triplet Loss-based Camera Network for ReID}},
  booktitle = {IJCNN},
  year = {2021},
  url = {https://openreview.net/forum?id=L99jCDL1ET},
}
Figure