Abstract
Person re-identification (re-ID) typically requires extensive cross-view labeling, which is labor-intensive and costly. This paper proposes One-View Learning (OVL), a practical paradigm where only a single camera view is labeled, while others remain unlabeled. The OVL framework addresses two primary challenges: the multi-target distribution gap across different cameras and the open-set problem of unknown identities in target views. By integrating Adversarial Multi-View Learning (AMVL) and Adversarial Unknown Rejection Learning (AURL), the proposed method achieves significant performance gains over source-only baselines and outperforms several unsupervised domain adaptation (UDA) techniques on major benchmarks like Market-1501 and DukeMTMC-reID.
Introduction
Person re-identification (re-ID) is the task of matching individuals across non-overlapping camera views. While deep learning has pushed the boundaries of re-ID performance, these models are traditionally "data-hungry," requiring exhaustive manual matching of the same person across dozens of cameras. This cross-view labeling process is the primary bottleneck for deploying re-ID systems in the real world.
The authors identify a "sweet spot" in the labeling process: it is much easier to label people within a single camera view (using automated tracking and detection) than it is to find matches across disjoint views. They define this as One-View Learning (OVL).
The Challenges of One-View Learning
Transitioning from fully supervised learning to OVL introduces two major technical hurdles:
- Multi-target Distribution Gap: Unlabeled images originate from multiple cameras, each with distinct lighting, angles, and backgrounds, creating a domain gap that hinders feature generalization.
- Open Set Problem: Target cameras often capture "unknown identities"—individuals who never appeared in the labeled source camera. Standard domain adaptation often fails here by incorrectly forcing these unknown samples to align with known source identities.
Methodology
The OVL framework utilizes a ResNet-50 backbone and two adversarial modules to overcome these hurdles.
Adversarial Multi-View Learning (AMVL)
To bridge the gap between cameras, the framework employs a multi-view classifier. During training, this classifier attempts to identify the source camera of an image, while the feature generator learns to produce view-invariant features that "deceive" the classifier. This adversarial process ensures the model focuses on identity-relevant features rather than camera-specific noise.
Adversarial Unknown Rejection Learning (AURL)
To address the open set problem, the authors introduce two identity classifiers designed to establish a decision boundary for "unknown" identities.
The classifiers attempt to push unlabeled target samples toward this "unknown" boundary. Simultaneously, the generator pushes samples either toward a known identity (if a match exists) or deep into the "unknown" category. This mechanism prevents the model from "polluting" the feature space with incorrect matches.
Experimental Results
The model was evaluated on the Market-1501, DukeMTMC-reID, and MSMT17 benchmarks. In each test, one camera was treated as the labeled source, with the remainder as unlabeled targets.
- Market-1501: OVL achieved 78.1% Rank-1 accuracy, a massive jump from the 49.8% baseline (source-only training).
- DukeMTMC-reID: Accuracy reached 58.5%, compared to the 33.9% baseline.
- Comparison: OVL outperformed established unsupervised domain adaptation (UDA) methods like HHL and SPGAN, despite the latter often using more complex data augmentation.
Conclusion
One-View Learning offers a compelling alternative to the high cost of full supervision. By explicitly modeling "unknown" identities and enforcing view invariance, the OVL framework provides a significant performance boost for a minimal increase in labeling effort. Future work will likely explore pseudo-labeling to further bridge the gap between OVL and fully supervised performance.
Summary updated