Person re-identification (Re-ID) is the task of recognizing the same individual across non-overlapping camera views. While early methods relied on simple visual descriptors, modern research focuses on learning representations that are invariant to significant changes in environment, sensor bias, and even the subject's own appearance. The goal is to distill a "pure" identity signature from a noisy visual signal.
Disentangling Identity from Transient Attributes
A primary hurdle in real-world Re-ID is that models often overfit to transient visual cues, such as clothing color or camera-specific lighting. When these cues change, traditional models fail. To address this, researchers have moved toward disentanglement—separating identity-stable features from identity-irrelevant ones.
In the context of cloth-changing Re-ID, Yan et al. proposed the UCAD framework to explicitly disentangle clothing features from identity. Similarly, Gao et al. introduced Identity-Guided Collaborative Learning (IGCL) to leverage human semantics to create clothing-invariant features. Rather than ignoring clothes, some methods use auxiliary information to guide the model: Trinh et al. utilize pose estimation to focus the encoder on body structure rather than fabric, while Yin et al. combine body-part attention with data augmentation to ensure the model looks past the attire.
This disentanglement also applies to environmental noise. Zhang et al. developed the CDE-Learning framework to eliminate "camera bias," using auxiliary tasks to identify and strip away noise specific to individual sensors, thereby improving the accuracy of unsupervised pseudo-labels.
Aggregating Information Across Space and Time
A single image is often an insufficient representation of identity due to occlusions or poor angles. Modern architectures improve robustness by aggregating information from multiple perspectives or time steps.
For video-based Re-ID, the challenge is to fuse temporal dynamics with spatial details. Liu et al. proposed Trigeminal Transformers (TMT) to simultaneously process spatial, temporal, and spatio-temporal views. This multi-view approach is echoed in image-based methods as well; Li et al. and Yu et al. demonstrated that using "multiple experts"—specialized models focusing on different feature subspaces—significantly boosts performance by combining diverse perspectives.
Spatial partitioning is another key mechanism. The SCR network, as described by the authors, uses a pyramidal structure to partition features both spatially and by channel, ensuring the model captures both fine-grained local details and global context. When parts of the body are hidden, Yong et al. show that "purifying" features through occlusion-aware transformers can prevent contaminated background noise from degrading the identity embedding.
Refining the Metric Space
Beyond feature extraction, the way we compare these features—the metric learning—has evolved. The intuition has shifted from comparing individual points to comparing distributions or sets.
Zhou et al. argued that "set-to-set" similarity is superior to point-to-point comparison because it accounts for the variance in a person's appearance across a sequence of images. By comparing sets, the model learns a more robust boundary between different individuals.
Loss functions have also become more specialized to enforce these boundaries. Wei et al. introduced Minimum Margin Loss (MML), which explicitly penalizes class centers that are too close together, forcing the model to learn more discriminative features. To handle the difficulty of training samples, Shi et al. proposed "moderate positive mining," which focuses the model on informative examples that are neither too easy nor impossibly hard, leading to better generalization in deep metric learning.
Cross-Modality and Domain Adaptation
Finally, the field is expanding toward scenarios where the source and target data look fundamentally different. This includes cross-view vehicle Re-ID, where Shang et al. used shape-based self-correlation to bridge the gap between aerial and ground-level perspectives.
In unsupervised domain adaptation, the goal is to transfer knowledge from a labeled dataset to an unlabeled one. Zhu et al. achieved this by using GANs to transfer styles between domains, allowing the model to learn target-specific features without manual labels. Similarly, Li et al. used reciprocal disentanglement to ensure that the features transferred across domains are strictly related to identity, rather than domain-specific artifacts.