The central challenge in modern surveillance and person search is the "supervision gap." While it is relatively easy to detect a pedestrian in a single frame, identifying that same individual across a network of non-overlapping cameras is a logistical nightmare for human annotators. To train a standard supervised model, a human must manually link thousands of identity pairs across different lighting conditions, angles, and occlusions—a process that Hong-Xing Yu and Ancong Wu describe as prohibitively expensive and a barrier to city-wide scaling in their work on CAMEL.
To bypass this bottleneck, researchers have developed a suite of techniques that allow models to learn identity representations from unlabeled or weakly labeled data. The core intuition is that the structure of the data itself—whether through visual similarity, temporal continuity in video, or cross-modal cues like captions—contains enough signal to bootstrap identity recognition.
Bootstrapping Identity through Iterative Clustering
The most direct way to handle unlabeled data is to let the model "guess" the labels and then refine those guesses. This is often framed as an iterative cycle: the model extracts features, clusters them to find likely identity groups, and then fine-tunes its feature extractor based on those clusters.
Hehe Fan and Liang Zheng formalized this in their Progressive Unsupervised Learning (PUL) framework. By alternating between $k$-means clustering and fine-tuning on only the most "reliable" samples, the model gradually learns to distinguish individuals without a single manual label. Similarly, Mang Ye and Andy J. Ma proposed Dynamic Graph Matching (DGM), which treats the problem as a graph where nodes are individuals and edges represent potential matches. As the model learns a better distance metric, the graph structure becomes clearer, effectively allowing the system to "self-supervise" its way to high performance.
Bridging the Gap with Synthetic and Source Data
Another powerful intuition is that identity features learned in one environment (the "source") should be somewhat useful in another (the "target"). This is the basis of Domain Adaptation.
Researchers like Junjie Li and Yichao Yan have extended this to Domain Adaptive Person Search (DAPS), where a model trained on a labeled dataset is adapted to a completely unlabeled target environment. To make this even more scalable, Qi Wang and Junyu Gao leverage high-fidelity synthetic data from video games like Grand Theft Auto V in their SMCR framework. By using style transfer to bridge the visual gap between the game and the real world, they provide the model with a "free" source of perfectly labeled identity data.
Generative Hallucination for Robustness
A major hurdle in re-identification is intra-class variation: the same person looks different because of their pose or the camera angle. Unsupervised models struggle here because they might cluster images based on "person in a red shirt facing left" rather than the identity itself.
To solve this, researchers use Generative Adversarial Networks (GANs) to "hallucinate" variations. Xuelin Qian introduced PN-GAN, which synthesizes images of a person in standardized "canonical" poses. By training on these pose-normalized images, the model learns to ignore the noise of body orientation. Similarly, Chengyuan Zhang used PAC-GAN to generate "pose-rich" synthetic samples, ensuring the model sees a person from every possible angle during training, even if the original unlabeled dataset is sparse.
Weak Supervision from Context and Text
Finally, when identity labels are missing, other forms of "weak" supervision can fill the void. In Weakly Supervised Person Search, Chuchu Han and colleagues show that simply having bounding boxes (knowing where a person is, but not who they are) is enough. Their Region Siamese Network (R-SiamNet) enforces consistency between the person's appearance and the surrounding scene context to anchor the identity.
This concept extends to cross-modal data. For instance, Derya Ozkan and Pinar Duygulu demonstrated that captions in news photos can act as noisy labels. By finding the "densest component" in a graph of face similarities associated with a specific name in a caption, the system can identify individuals across thousands of "in-the-wild" photos without any manual training. Similarly, Changxin Gao uses Cross-Modal Mutual Training to link text descriptions to images, leveraging noun-phrase similarity to refine identity clusters.