Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Source-Guided Similarity Preservation for Online Person Re-Identification

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

TL;DR

The paper tackles Online Unsupervised Domain Adaptation for person Re-Identification (OUDA-Rid), focusing on catastrophic forgetting and domain shift when learning from a streaming unlabeled target. It introduces Source-Guided Similarity Preservation (S2P), which constructs a source-based support set to regularize learning via a teacher-student knowledge-distillation mechanism and explicit distribution alignment, all without storing target data to satisfy privacy constraints. S2P is designed as a modular framework that can incorporate existing pseudo-labeling approaches (MMT, SpCL, IDM) and demonstrates substantial performance gains on real-to-real and synthetic-to-real OUDA benchmarks. Ablation studies validate the effectiveness of the KD and MMD losses, the similarity-based support selection, and the EMA teacher, supporting the method’s practical impact for privacy-conscious continual Re-ID in evolving environments.

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.

Source-Guided Similarity Preservation for Online Person Re-Identification

TL;DR

The paper tackles Online Unsupervised Domain Adaptation for person Re-Identification (OUDA-Rid), focusing on catastrophic forgetting and domain shift when learning from a streaming unlabeled target. It introduces Source-Guided Similarity Preservation (S2P), which constructs a source-based support set to regularize learning via a teacher-student knowledge-distillation mechanism and explicit distribution alignment, all without storing target data to satisfy privacy constraints. S2P is designed as a modular framework that can incorporate existing pseudo-labeling approaches (MMT, SpCL, IDM) and demonstrates substantial performance gains on real-to-real and synthetic-to-real OUDA benchmarks. Ablation studies validate the effectiveness of the KD and MMD losses, the similarity-based support selection, and the EMA teacher, supporting the method’s practical impact for privacy-conscious continual Re-ID in evolving environments.

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.
Paper Structure (11 sections, 8 equations, 4 figures, 4 tables)

This paper contains 11 sections, 8 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Source-Guided Similarity Preservation
  4. Overview of the Approach
  5. Source-Guided Knowledge Distillation
  6. Source-Target Distribution Alignment
  7. Incorporating Pseudo-Labeling into S2P.
  8. Experiments and Results
  9. Quantitative Results
  10. Ablation Studies
  11. Conclusions

Figures (4)

  • Figure 1: In OUDA for person Re-ID, the images of the target domain are available as a stream of data, and past images cannot be stored. Two main challenges should be addressed: 1) catastrophic forgetting and 2) domain shift.
  • Figure 2: The pipeline of S2P. a) S2P incorporates knowledge distillation $\mathcal{L}_{KD}$, discrepancy $\mathcal{L}_{MMD}$ loss functions, and a teacher model to mitigate the catastrophic forgetting and domain-shift problems. b) Our algorithm employs a similarity-based selection to construct the support set $\xi_k$ from the source domain that maximizes the similarity with the target images.
  • Figure 3: Comparison of S2P with four state-of-the-art methods in terms of mAP vs. task index in two different OUDA-Rid tasks, MSMT$\to$CUHK and RandPerson$\to$Market.
  • Figure 4: The support set construction based on the similarities between the source domain MS (RP respectively) and the target domain C (M respectively).