Source-Guided Similarity Preservation for Online Person Re-Identification

Hamza Rami, Jhony H. Giraldo, Nicolas Winckler, Stéphane Lathuilière

Abstract

Online Unsupervised Domain Adaptation (OUDA) for person Re-Identification (Re-ID) is the task of continuously adapting a model trained on a well-annotated source-domain dataset to a target domain observed as a data stream. In OUDA, person Re-ID models face two main challenges: catastrophic forgetting and domain shift. In this work, we propose a new Source-guided Similarity Preservation (S2P) framework to alleviate these two problems. Our framework is based on the extraction of a support set composed of source images that maximizes the similarity with the target data. This support set is used to identify feature similarities that must be preserved during the learning process. S2P can incorporate multiple existing UDA methods to mitigate catastrophic forgetting. Our experiments show that S2P outperforms previous state-of-the-art methods on multiple real-to-real and synthetic-to-real challenging OUDA benchmarks.

Introduction: The Challenges of Online Re-ID

Person Re-Identification (Re-ID) involves recognizing the same individual across non-overlapping camera views. While modern deep learning models achieve high accuracy on static benchmarks, they often struggle with "domain shift" when deployed in new environments. Standard Unsupervised Domain Adaptation (UDA) addresses this by training on target data, but real-world deployment introduces Online Unsupervised Domain Adaptation (OUDA) constraints.

Figure 4: Challenges in Online Unsupervised Domain Adaptation (OUDA), including domain shift between source and target domains and catastrophic forgetting within the target data stream.

OUDA presents two primary hurdles:

  1. Privacy and Online Constraints: Data arrives as a continuous stream. Privacy regulations like GDPR often prohibit storing target images, preventing the use of "replay buffers" to mitigate forgetting.
  2. Catastrophic Forgetting: As models adapt to new stream data, they lose discriminative features learned from the source domain or previous stream segments.

The S2P Framework

The proposed Source-guided Similarity Preservation (S2P) framework addresses these challenges by using the source domain as a privacy-safe proxy for memory.

Figure 6: (a) Overall pipeline of the S2P framework featuring Student and Teacher models with similarity-based selection. (b) Detail of the support set construction via similarity-based selection in the feature space.

The framework consists of three core components:

  • Source-Guided Support Set: Instead of storing restricted target images, S2P selects a subset of source images (which are permissible to store) that are visually similar to the current target data. This support set approximates the target domain's feature space.
  • Similarity-Based Knowledge Distillation (KD): S2P employs a distillation loss that preserves relative distances (similarities) between images. A "Teacher" model (an exponential moving average of the student) ensures that structural relationships between identities remain stable during adaptation.
  • Distribution Alignment (MMD): Maximum Mean Discrepancy (MMD) loss is used to minimize the distance between source and target feature distributions, enhancing domain invariance.

Experimental Evaluation

The authors evaluated S2P across real-to-real and synthetic-to-real benchmarks using datasets such as Market-1501 (M), MSMT17 (MS), CUHK03 (C), and RandPerson (RP). Target datasets were split into five sequential tasks to simulate an online stream.

Key Results

S2P consistently outperformed state-of-the-art methods in the online setting:

  • Performance Gains: In the MS → C task, S2P-MMT achieved 40.4% mAP, significantly higher than the baseline MMT's 32.2%.
  • Synthetic-to-Real Adaptation: On the RP → M task, S2P-IDM reached 70.2% mAP, compared to 60.8% for the baseline.
  • Stability: Unlike baseline models that exhibited performance dips due to forgetting, S2P maintained a steady upward trend in accuracy throughout the data stream.

Conclusion

S2P provides a robust, privacy-compliant solution for online person Re-ID. By leveraging source data as a proxy for memory and focusing on preserving the structural similarity of the feature space, it effectively mitigates catastrophic forgetting and domain shift. Future work may explore more complex streams with varying environmental conditions.

Summary updated

Abstract

Online Unsupervised Domain Adaptation (OUDA) for person Re-Identification (Re-ID) is the task of continuously adapting a model trained on a well-annotated source-domain dataset to a target domain observed as a data stream. In OUDA, person Re-ID models face two main challenges: catastrophic forgetting and domain shift. In this work, we propose a new Source-guided Similarity Preservation (S2P) framework to alleviate these two problems. Our framework is based on the extraction of a support set composed of source images that maximizes the similarity with the target data. This support set is used to identify feature similarities that must be preserved during the learning process. S2P can incorporate multiple existing UDA methods to mitigate catastrophic forgetting. Our experiments show that S2P outperforms previous state-of-the-art methods on multiple real-to-real and synthetic-to-real challenging OUDA benchmarks.
BibTeX
@inproceedings{rami2024source,
  author = {Hamza Rami and Jhony H. Giraldo and Nicolas Winckler and Stéphane Lathuilière},
  title = {{Source-Guided Similarity Preservation for Online Person Re-Identification}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2024},
  url = {https://openreview.net/forum?id=Nlt25YP3AF},
}
Figure