MSBA: Multiple Scales, Branches and Attention Network With Bag of Tricks for Person Re-Identification

Hanlin Tan, Huaxin Xiao, Xiaoyu Zhang, Bin Dai, Shiming Lai, Yu Liu, Maojun Zhang

Abstract

The MSBA (Multiple Scales, Branches, and Attention) network achieves state-of-the-art performance in person re-identification (Re-ID) by integrating a robust multi-component architecture with a refined "bag of tricks" for training. Beyond architectural novelty, this research highlights how specific training strategies—such as a custom sampling method for unbalanced data—are often more critical for scaling to massive, real-world datasets than the model design itself.

Introduction

Person Re-ID—the task of identifying the same individual across different non-overlapping camera views—is plagued by variations in lighting, pose, occlusion, and image quality. While many recent models boast high accuracy on small, curated public datasets, they often fail to generalize when deployed on "super big" private datasets or real-world scenarios.

The authors identify a growing gap: many academic designs are over-fitted to the high-quality images of small benchmarks. Furthermore, some high-performing models are "industry-unfriendly" due to massive parameter counts and slow inference times. The MSBA model is designed to be both highly accurate and robust enough for large-scale deployment.

Methodology

The proposed MSBA architecture is built on a ResNet-50-IBN backbone. The "IBN" (Instance-Batch Normalization) component is crucial, as it helps the model generalize across different image contrasts and lighting conditions without adding computational overhead.

The architecture consists of three key pillars:

  1. Multi-Scale Branch: This extracts features from early stages (Stages 2 and 3) of the backbone. This helps the model capture details at different granularities, which is vital when people appear at varying distances from the camera.
  2. Multi-Branch (Global & Local): The model uses a global branch for the whole body and a local branch that "slices" the feature map into upper and lower halves. This forces the model to learn specific features for different body parts, aiding in cases of partial occlusion.
  3. Attention Mechanism: The authors integrate Squeeze-and-Excitation (SE) modules to perform channel-wise attention. Interestingly, they also implement "inverse attention" as a form of regularization, ensuring that features discarded by the attention modules are still constrained by an identity (ID) loss.

The training process utilizes a combination of Triplet Loss (for metric learning) and Cross-Entropy Loss (for classification).

Training Strategies and "Bag of Tricks"

One of the paper's most significant takeaways is that training "tricks" can be just as important as the model architecture itself. Key strategies include:

  • Warm-up Learning Rates: Gradually increasing the learning rate to stabilize early training.
  • Random Erasing: A data augmentation technique to improve robustness against occlusion.
  • Label Smoothing: Preventing the model from becoming over-confident in its classifications.
  • Smart Sampling for Unbalanced Data: Real-world data is often "long-tailed" (some individuals appear in hundreds of images, others in only a few). To prevent over-fitting to frequently seen individuals, the authors introduced a sampling threshold ($\sigma$) that restricts the maximum number of instances per person ID used in each training epoch.

Experimental Results

The authors evaluated MSBA on four major public benchmarks without using common post-processing "cheats" like re-ranking or flipping inference:

  • Market1501: Achieved a Rank-1 accuracy of 95.8% and an mAP of 89.0%.
  • MSMT17: On this massive and challenging dataset, MSBA reached a Rank-1 of 76.1% and an mAP of 60.2%, significantly outperforming previous benchmarks.
  • DukeMTMC-ReID & CUHK03: The model showed consistent state-of-the-art performance across medium-sized and automatically detected (misaligned) datasets.

Ablation studies confirmed the necessity of each component; for instance, removing the attention modules caused a 2.1-point drop in the final score on the CUHK03 dataset.

Discussion and Conclusion

MSBA demonstrates that high performance in Re-ID does not require overly complex part-based slicing, which often generalizes poorly to large datasets due to alignment errors. Instead, a combination of multi-scale features, attention, and robust training strategies provides a more scalable solution.

Furthermore, MSBA maintains inference efficiency. The extra branches used for multi-scale features are removed during the inference stage, allowing the model to benefit from multi-scale regularization during training without a computational penalty during deployment. Future work may explore unsupervised learning strategies to leverage the vast amounts of unlabeled data available in real-world surveillance systems.

Summary updated

Abstract

Person re-identification (Re-ID) has become a hot topic in both research and industry. We joined in a person Re-ID challenge of the First National Artificial Intelligence Challenge (China, 2019) and found some model designs and training tricks work great or not on a super big private dataset. In this paper, we propose a model that combines the most effective designs, including multi-scale, multi-branch and attention mechanism, and report training tricks that are no less or even more important in improving person Re-ID performance. We analyze four commonly used public datasets: Market1501, DukeMTMC-ReID, CUHK03, and MSMT17, and achieve the state-of-the-art performance. Besides, we analyze and confirm the effectiveness of the designs by ablation studies. We also share strategies that play a key role in the challenge and experience of model designs that do not generalize well on large datasets.
BibTeX
@article{tan2020msba,
  author = {Hanlin Tan and Huaxin Xiao and Xiaoyu Zhang and Bin Dai and Shiming Lai and Yu Liu and Maojun Zhang},
  title = {{MSBA: Multiple Scales, Branches and Attention Network With Bag of Tricks for Person Re-Identification}},
  journal = {IEEE Access},
  year = {2020},
  doi = {10.1109/access.2020.2984915},
}
Figure