Abstract
Recently, GAN-based methods have demonstrated strong effectiveness in generating augmentation data for person re-identification (ReID), bridging the gap between domains and enriching data variety. However, most ReID works utilize all GAN-generated data indiscriminately or evaluate quality at the dataset level, ignoring image-level features essential for ReID. This paper analyzes the characteristics of ReID samples to determine what makes a GAN-generated image "good." We propose examining each sample with id-consistency and diversity constraints by mapping images into different feature spaces. Using a metric-based sampling method, we demonstrate that curated synthetic data significantly outperforms full augmentation sets. Our method shows effectiveness across supervised and unsupervised domain adaptation tasks.
Introduction: The Problem with "Black Box" Augmentation
Person Re-Identification (ReID) involves matching individuals across non-overlapping camera views, a task complicated by variations in lighting, pose, and camera style. While Generative Adversarial Networks (GANs) are frequently used to expand training sets, most existing methods treat GAN output as a black box. They often use all generated images or rely on dataset-level metrics like FID, which can include noisy images that lose identity information or redundant images that add no value. This research seeks to define and select high-quality synthetic samples that actually benefit the ReID model.
Methodology: Metric-based Sampling
The proposed pipeline filters synthetic data through two distinct feature spaces before training:
- Consistency Space ($E_c$): Captures identity-related features (clothing, hair, gender) while ignoring style. An image is kept if it is close to the "identity center" of real images.
- Diversity Space ($E_d$): Captures variations like pose and lighting. An image is kept if it is sufficiently far from real images, ensuring it provides new information.
Additionally, the authors use Local Outlier Factor (LOF) density filtering to remove clusters of nearly identical synthetic images, preventing overfitting on GAN artifacts. Finally, a refined training strategy is employed: while real data uses both Cross-Entropy and Triplet loss, filtered GAN data uses only Cross-Entropy loss to mitigate sensitivity to remaining noise.
Key Insights: Quality Over Quantity
The study reveals that blindly adding GAN data can degrade performance. On the MSMT-17 dataset, using the full GAN output dropped Rank@1 accuracy, whereas using the filtered "good" images improved it.
- The "Goldilocks" Zone: Ideal synthetic images must be consistent enough to maintain identity but diverse enough to teach the model new environmental variations.
- Disentanglement and Diversity Proxies: Using specialized encoders (like ISGAN) for identity and general models (like ImageNet-pretrained ResNet-50) for diversity proved most effective. Surprisingly, ImageNet-trained models are better judges of environmental diversity than models trained specifically on humans.
Results and Performance
The proposed method consistently outperformed state-of-the-art benchmarks:
- Supervised ReID (Market-1501): Achieved 96.2% Rank@1 and 89.2% mAP, surpassing both the baseline and full-augmentation approaches.
- Unsupervised Domain Adaptation (Duke → Market): The baseline mAP of 62.8% was boosted to 78.6%, outperforming specialized UDA methods like MMT.
- Efficiency: The system achieved superior results using only approximately 30% of the total generated GAN data, highlighting the power of selective sampling.
Conclusion
This work demonstrates that not every GAN-generated image is beneficial for ReID. By applying identity-consistency and diversity constraints, researchers can significantly improve model robustness and accuracy while reducing the computational load of training on redundant or noisy synthetic data. Future work may explore more sophisticated sampling metrics and improved disentanglement encoders to further refine the selection process.
Summary updated