Mixed High-Order Attention Network for Person Re-Identification

Binghui Chen, Weihong Deng, Jiani HuICCV 2019 · OpenReview

Abstract

Attention has become more attractive in person reidentification (ReID) as it is capable of biasing the allocation of available resources towards the most informative parts of an input signal. However, state-of-the-art works concentrate only on coarse or first-order attention design, e.g. spatial and channels attention, while rarely exploring higher-order attention mechanism. We take a step towards addressing this problem. In this paper, we first propose the High-Order Attention (HOA) module to model and utilize the complex and high-order statistics information in attention mechanism, so as to capture the subtle differences among pedestrians and to produce the discriminative attention proposals. Then, rethinking person ReID as a zero-shot learning problem, we propose the Mixed High-Order Attention Network (MHN) to further enhance the discrimination and richness of attention knowledge in an explicit manner. Extensive experiments have been conducted to validate the superiority of our MHN for person ReID over a wide variety of state-of-the-art methods on three large-scale datasets, including Market-1501, DukeMTMC-ReID and CUHK03-NP.

Introduction

Person Re-Identification (ReID) involves matching individuals across disjoint camera views, a task complicated by variations in pose, lighting, and occlusion. While attention mechanisms have significantly improved ReID by focusing on salient regions, most existing methods rely on "coarse" first-order attention. Spatial attention identifies where to look, and channel attention determines what features are important.

However, these first-order methods often overlook the complex, fine-grained interactions between visual elements that are crucial for distinguishing between similar-looking individuals. For instance, identifying a specific person might require recognizing the relationship between a backpack strap's texture and the color of a shirt—a high-order correlation that standard attention maps might miss.

Comparison of spatial, channel, and high-order attention mechanisms.

Methodology

High-Order Attention (HOA) Module

The core innovation is the High-Order Attention (HOA) module. Unlike standard attention that uses simple scaling factors, the HOA module employs a linear polynomial predictor to model high-order statistics of convolutional activations. To manage the computational complexity and parameter explosion typically associated with high-order tensors, the authors utilize Tensor Decomposition. By approximating high-order tensors with multiple rank-1 tensors, the HOA module captures intricate feature interactions without becoming computationally bloated.

Mixed High-Order Attention Network (MHN)

The authors propose the Mixed High-Order Attention Network (MHN), which integrates multiple HOA modules of varying orders (e.g., from 1st up to 6th order) into a unified architecture. This multi-branch approach ensures that the model captures a diverse range of features, from coarse global patterns to highly specific local interactions.

The Mixed High-Order Attention Network (MHN) architecture.

A significant challenge in multi-order networks is "order collapse," where high-order branches default to learning simpler, lower-order patterns. To mitigate this, the MHN incorporates an Adversarial Learning Constraint. A discriminator is trained to make features from different branches indistinguishable, while the HOA branches are trained to maximize their distinctiveness. This adversarial process forces each branch to focus on unique, order-specific information.

ReID as Zero-Shot Learning

The paper re-evaluates ReID through the lens of Zero-Shot Learning (ZSL). Since the identities encountered during testing are never seen during training, models must learn highly generalizable and rich feature representations. By explicitly modeling high-order statistics, the MHN provides a more comprehensive set of descriptors that are less likely to be biased toward the specific training set, improving generalization to unseen pedestrians.

Results and Discussion

The MHN-6 (a 6-branch configuration) was evaluated on three major benchmarks: Market-1501, DukeMTMC-ReID, and CUHK03-NP. It consistently outperformed baseline architectures (IDE and PCB) and current state-of-the-art (SOTA) attention-based methods:

  • Market-1501: Achieved a Rank-1 accuracy of 95.1% and an mAP of 85.0%, a substantial jump over the PCB baseline.
  • DukeMTMC-ReID: Reached 89.1% Rank-1 and 77.2% mAP.
  • CUHK03-NP: On the labeled set, it achieved 77.2% Rank-1, outperforming previous SOTA methods like CASN+PCB by nearly 4%.

Ablation studies confirmed that both high-order modeling and the adversarial constraint are vital for these performance gains. While increasing the number of branches ($k$) improves results, the authors noted diminishing returns beyond $k=6$, suggesting a point of optimal complexity for current datasets.

Conclusion

The Mixed High-Order Attention Network represents a shift from coarse to fine-grained attention in person ReID. By leveraging high-order statistics and an adversarial training regime, the MHN captures the subtle cues necessary for robust cross-camera matching. This approach not only sets new performance benchmarks but also provides a more robust framework for addressing the zero-shot nature of the ReID task.

Summary updated

Abstract

Attention has become more attractive in person re-identification (ReID) as it is capable of biasing the allocation of available resources towards the most informative parts of an input signal. However, state-of-the-art works concentrate only on coarse or first-order attention design, e.g. spatial and channels attention, while rarely exploring higher-order attention mechanism. We take a step towards addressing this problem. In this paper, we first propose the High-Order Attention (HOA) module to model and utilize the complex and high-order statistics information in attention mechanism, so as to capture the subtle differences among pedestrians and to produce the discriminative attention proposals. Then, rethinking person ReID as a zero-shot learning problem, we propose the Mixed High-Order Attention Network (MHN) to further enhance the discrimination and richness of attention knowledge in an explicit manner. Extensive experiments have been conducted to validate the superiority of our MHN for person ReID over a wide variety of state-of-the-art methods on three large-scale datasets, including Market-1501, DukeMTMC-ReID and CUHK03-NP. Code is available at http://www.bhchen.cn.
BibTeX
@inproceedings{chen2019mixed,
  author = {Binghui Chen and Weihong Deng and Jiani Hu},
  title = {{Mixed High-Order Attention Network for Person Re-Identification}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2019},
  url = {https://openreview.net/forum?id=_xkLVF3FUzd},
}
Figure