Abstract
Recently, person re-identification (ReID) has vastly benefited from the surging waves of data-driven methods. However, these methods are still not reliable enough for real-world deployments, due to the insufficient generalization capability of the models learned on existing benchmarks that have limitations in multiple aspects, including limited data scale, capture condition variations, and appearance diversities. To this end, we collect a new dataset named Person30K with the following distinct features: 1) a very large scale containing 1.38 million images of 30K identities, 2) a large capture system containing 6,497 cameras deployed at 89 different sites, 3) abundant sample diversities including varied backgrounds and diverse person poses. Furthermore, we propose a domain generalization ReID method, dual-meta generalization network (DMG-Net), to exploit the merits of meta-learning in both the training procedure and the metric space learning. Concretely, we design a "learning then generalization evaluation" meta-training procedure and a meta-discrimination loss to enhance model generalization and discrimination capabilities. Comprehensive experiments validate the effectiveness of our DMG-Net.
Introduction
Person Re-identification (ReID) is the task of matching the same individual across different, non-overlapping camera views. While deep learning has pushed ReID performance to high levels on specific benchmarks, these models often "break" when deployed in the real world. The authors identify two main culprits: existing datasets are too small (usually ~4K identities and <20 cameras) and lack environmental diversity, and models tend to overfit to the specific "style" of their training cameras rather than learning identity-invariant features.
The Person30K Dataset
To address these limitations, the authors introduce Person30K, a massive new dataset that is 11 times larger than previous benchmarks. It features 1.38 million images of 30,000 identities, collected from 6,497 cameras across 89 different sites (supermarkets, malls, streets) over a full year.
As shown in Figure 1, Person30K provides a much broader range of backgrounds and person poses than traditional datasets. The dataset includes detailed annotations for person ID, camera ID, site ID, and scenario category.
Dual-Meta Generalization Network (DMG-Net)
To handle this scale and improve real-world performance, the authors propose DMG-Net, a dual-meta learning framework that simulates "unseen" camera environments during training. The "dual" refers to two distinct meta-learning applications:
- Meta-Generalization Training: Instead of standard training, DMG-Net uses a "learning then generalization evaluation" loop. In each mini-batch, data is split into a support set and a query set from non-overlapping cameras. The model learns on the support set and then evaluates itself on the query set, updating its weights specifically to perform better on "unseen" cameras.
- Meta-Discrimination Loss: To improve the metric space, the model explicitly mimics the cross-camera matching process. It calculates feature-based centers for identities in the support set and forces query samples of the same identity to cluster near these centers while pushing different identities away.
Key Results and Insights
- Scale is a Teacher: Training on Person30K significantly boosts a model's ability to generalize. A model trained on Person30K achieves 71.1% mAP on Market1501—a massive leap over models trained on other existing benchmarks without any fine-tuning.
- Simulating the "Unseen": By constantly forcing the model to "test" itself on different cameras during every training step, DMG-Net learns to prioritize features (like clothing patterns or body shape) that remain invariant across different viewpoints.
- State-of-the-Art Performance: DMG-Net outperformed existing methods across all subsets of Person30K. On the most difficult "Test-C" (different scenario) subset, it achieved 64.39% mAP.
Conclusion
Person30K and DMG-Net represent a significant advancement in creating robust ReID systems for real-world deployment. While meta-learning involves higher computational costs, the resulting generalization capabilities are essential for practical applications. Future work may explore the potential of Person30K for unsupervised or semi-supervised domain adaptation.
Summary updated