VIPerson: Flexibly Generating Virtual Identity for Person Re-Identification

Xiao-Wen Zhang, Delong Zhang, Yi-Xing Peng, Zhi Ouyang, Jingke Meng, Wei-Shi ZhengICCV 2025

TL;DR

This paper introduces VIPerson, a novel pipeline for generating realistic synthetic pedestrian images with flexible virtual identities for person re-identification (ReID). By addressing key challenges like identity diversity, cross-camera variations, and camera-realistic styles, VIPerson aims to overcome the limitations of real-world ReID datasets, such as high labeling costs and privacy concerns, while significantly improving ReID model performance.

What Problem Does This Paper Address?

Person re-identification (ReID) is crucial for tasks like surveillance and security, aiming to match individuals across different camera views. However, training effective ReID models requires vast amounts of labeled real-world data, which is expensive to collect and annotate. Furthermore, using real-world data raises significant privacy concerns due to the sensitive nature of pedestrian images, leading to strict regulations and even the removal of public datasets. Existing synthetic data generation methods fall short: engine-based methods lack realism, while learning-based methods often entangle synthetic styles with real identities, posing privacy risks. They also struggle with generating diverse identities and sufficient cross-camera variations. This paper tackles these issues by proposing a method to generate high-quality, privacy-preserving synthetic ReID data with flexible control over identity, pose, scene, and style.

Methods

VIPerson employs a three-pronged approach to generate synthetic pedestrian images:

  • Flexible Identity Generator: This component builds upon a latent diffusion model (LDM) and enhances it with a "dropout text embedding" technique. This allows for the flexible generation of both new random identities and "hard identities" – individuals that share subtle visual similarities, making them challenging for ReID models. The model is trained on a large dataset of virtual identities to learn a rich distribution.
  • Cross-Camera Diversification: To ensure generated images capture realistic variations, VIPerson introduces scalable cross-camera variations. This involves:
  • Pose Variations: Using the IP-adapter, generated identities are adapted to various poses extracted from real-world images.
  • Scene Variations: A "segment-and-paste" strategy is used to replace the original scene with new ones, creating diverse environmental contexts for each identity.
  • Style Realization: To bridge the realism gap between synthetic and real images, a neural style transfer method is applied. This technique injects camera-realistic styles (e.g., illumination, motion blur) from real-world reference images into the generated synthetic images in an identity-agnostic manner, preserving the virtual identity while enhancing realism.

The pipeline leverages the large-scale unlabeled LUPerson dataset for collecting diverse poses, scenes, and styles.

Datasets

  • Training Data: The VIPerson pipeline is trained using a large-scale virtual identity dataset called MALS (1.5 million images) and the LUPerson dataset (over 4.18 million images) for pose, scene, and style information.
  • Evaluation Datasets: The generated VIPerson dataset is evaluated on downstream ReID tasks using real-world datasets like MSMT17 and Market1501.
  • New Benchmark: The paper introduces the VIPerson dataset itself, which is composed entirely of virtual identities and is designed to be flexible in terms of identity diversity, cross-camera variations, and camera-realistic style. They also propose VIPerson*, which incorporates scenario-specific information for targeted applications.

Key Insights and Ideas

  • The Power of "Hard Identities": Generating "hard identities" (individuals with subtle visual similarities) significantly boosts ReID model performance, forcing models to learn finer-grained discriminative features. This is a key advantage over previous synthetic methods that often struggle to create such challenging samples.
  • Scalability is Key: The proposed methods for generating diverse identities, cross-camera variations (poses and scenes), and realistic styles are designed to be scalable, allowing for the creation of large-scale datasets without manual annotation bottlenecks.
  • Identity-Agnostic Style Transfer is Crucial: By decoupling style from identity during style transfer, VIPerson avoids the privacy pitfalls of learning-based methods that embed real identity information into synthetic styles. This ensures privacy while achieving high realism.
  • Synthetic Data Can Outperform Real Data (in some scenarios): The results demonstrate that VIPerson can achieve superior performance in direct transfer and unsupervised domain adaptation tasks compared to existing synthetic datasets and even sometimes rivaling real-world pre-training, highlighting its potential to overcome real-world data limitations.

Key Results

  • Direct Transfer: VIPerson significantly outperforms existing synthetic datasets on direct transfer tasks on MSMT17 and Market1501, achieving rank-1 accuracies of 32.5% and 66.5% respectively.
  • Supervised Fine-tuning: Pre-training with VIPerson and fine-tuning on target datasets leads to state-of-the-art performance, reaching 76.7% rank-1 on MSMT17 and 95.1% on Market1501.
  • Unsupervised Domain Adaptation: VIPerson-based pre-training achieves strong results in unsupervised domain adaptation, outperforming other synthetic datasets and demonstrating its effectiveness without manual annotations.
  • Identity Expansion: In scenarios with very few labeled identities, augmenting with VIPerson significantly improves ReID performance, showcasing its utility in privacy-sensitive applications.

Limitations and Future Work

  • Performance Bottleneck: The authors observe a performance bottleneck at a certain data scale, suggesting that future work should focus on how to effectively leverage synthetic data to enhance ReID models, rather than solely increasing data scale.
  • Future Directions: The paper suggests exploring how to further leverage synthetic data to enhance ReID models and how to effectively incorporate scenario-specific patterns for targeted applications. The flexibility of the VIPerson pipeline also opens avenues for exploring its use in other vision tasks beyond ReID.

Summary updated

BibTeX
@inproceedings{zhang2025viperson,
  author = {Xiao-Wen Zhang and Delong Zhang and Yi-Xing Peng and Zhi Ouyang and Jingke Meng and Wei-Shi Zheng},
  title = {{VIPerson: Flexibly Generating Virtual Identity for Person Re-Identification}},
  booktitle = {IEEE/CVF International Conference on Computer Vision},
  year = {2025},
}
Figure