A Strong Baseline and Batch Normalization Neck for Deep Person Re-identification (1906.08332v2)

Abstract: This study explores a simple but strong baseline for person re-identification (ReID). Person ReID with deep neural networks has progressed and achieved high performance in recent years. However, many state-of-the-art methods design complex network structures and concatenate multi-branch features. In the literature, some effective training tricks briefly appear in several papers or source codes. The present study collects and evaluates these effective training tricks in person ReID. By combining these tricks, the model achieves 94.5% rank-1 and 85.9% mean average precision on Market1501 with only using the global features of ResNet50. The performance surpasses all existing global- and part-based baselines in person ReID. We propose a novel neck structure named as batch normalization neck (BNNeck). BNNeck adds a batch normalization layer after global pooling layer to separate metric and classification losses into two different feature spaces because we observe they are inconsistent in one embedding space. Extended experiments show that BNNeck can boost the baseline, and our baseline can improve the performance of existing state-of-the-art methods. Our codes and models are available at: https://github.com/michuanhaohao/reid-strong-baseline.

• The paper demonstrates the efficacy of combining training techniques like warm-up, random erasing, label smoothing, and BNNeck to boost model performance.
• It introduces BNNeck to separate metric and classification features, yielding 94.5% rank-1 accuracy and 85.9% mAP on Market1501.
• The study emphasizes intra-class compactness and inter-class separability, offering a streamlined alternative to complex multi-branch networks.

A Strong Baseline and Batch Normalization Neck for Deep Person Re-identification

Overview

The paper introduces a robust baseline for deep person re-identification (ReID), emphasizing simplicity and effectiveness. Existing state-of-the-art methods often involve complex networks with multiple branches, which increases computational overhead. In contrast, this work amalgamates effective training techniques scattered across various sources, achieving remarkable performance using only ResNet50 global features. The baseline attains 94.5% rank-1 accuracy and 85.9% mAP on Market1501, outperforming established global- and part-based baselines.

Key Contributions

1. Training Tricks: The authors consolidate and evaluate several training tricks that enhance model performance, including warm-up learning rates, random erasing augmentation, label smoothing, altered last stride, and the introduction of a batch normalization neck (BNNeck).
2. BNNeck Architecture: A novel contribution is the BNNeck, which adds a BN layer post-global pooling, effectively separating metric and classification losses into distinct feature spaces. This separation addresses the inconsistency observed when both losses share a unified embedding space. Experiments demonstrate that BNNeck enhances the baseline, often outperforming existing methods.
3. Intra-Class Compactness and Inter-Class Separability: The authors highlight the importance of clustering characteristics in ReID tasks, often overshadowed by a sole focus on ranking metrics. They propose utilizing center loss to improve clustering effects, which is crucial for tasks like object tracking that rely on distance thresholds for classification.
4. Experimentation and Verification: Extended experiments validate the general effectiveness of the proposed tricks across various domains and network backbones, confirming the robustness and adaptability of the approach.

Implications

The implications of this research extend to both practical applications and theoretical understandings of ReID:

• Practical Impact: The streamlined design and implementation of a strong baseline model sans additional computation costs are particularly beneficial for real-world applications where simplicity and efficiency are paramount, such as video surveillance and criminal investigations.
• Theoretical Insights: By dissecting the feature space into separate avenues for classification and metric learning, the paper offers insights into optimizing model architecture for enhanced performance without resorting to overly complex multi-branch strategies.

Future Directions

The research suggests promising directions for extending person ReID capabilities. With an emphasis on simplicity combined with high performance, future work could explore additional training techniques or network architectures that can further elevate the efficacy of ReID systems while maintaining or reducing computational overhead. Moreover, the noted discrepancy in cross-domain performance when utilizing certain data augmentation strategies, such as random erasing, invites further investigation.

In summary, the paper contributes a significant stride toward efficient and effective person re-identification through a consolidated approach of training enhancements and architectural innovations, facilitating advances in both academic research and industrial application of ReID systems.