Meta Distribution Alignment for Generalizable Person Re-Identification

Hao Ni, Jingkuan Song, Xiaopeng Luo, Feng Zheng, Wen Li, Heng Tao Shen

Abstract

Person Re-Identification (ReID) models often suffer from performance degradation when deployed in unseen environments due to the domain gap. While Unsupervised Domain Adaptation (UDA) offers a solution, it requires target-domain data during training. This paper introduces Meta Distribution Alignment (MDA), a framework for Domain Generalizable (DG) ReID. MDA aligns features from multiple source domains to a universal prior distribution, creating a "bridge" to unseen target domains. Combined with a meta-learning strategy and test-time adaptation, MDA achieves state-of-the-art results on major benchmarks.

Introduction

The goal of Person Re-Identification (ReID) is to match individuals across non-overlapping camera views. While deep learning has pushed performance to high levels within closed datasets, real-world deployment often involves "unseen" locations where no training data exists. This challenge, termed Domain Generalizable (DG) ReID, requires models to learn features that are robust to environmental shifts without prior exposure to the target domain.

Conceptual schematic of Meta Distribution Alignment (MDA) showing the alignment of source and target features to a universal prior distribution.

The core difficulty in DG ReID is aligning source features with an unknown target distribution. The authors propose using a universal prior distribution as an intermediary, effectively bridging the gap between seen and unseen domains.

Methodology

The MDA framework integrates three primary components within a meta-learning paradigm to enhance feature robustness and adaptability.

Detailed architecture of the MDA framework, illustrating the ID feature learning, prior distribution alignment, and distribution-guided refining modules.

1. Prior Distribution Alignment

Since the target domain is inaccessible during training, MDA aligns source features to a known prior distribution (e.g., a standard Gaussian). High-dimensional identity features are mapped into a low-dimensional latent space using an encoder-decoder structure. The Wasserstein distance is employed to minimize the discrepancy between these latent embeddings and the prior. This alignment serves as a "universal bridge": if multiple source domains are close to the prior, they are implicitly aligned with each other and, theoretically, with an unseen target domain that also undergoes alignment.

2. Distribution Guided Refining

Strictly constraining features to a simple distribution can lead to the loss of discriminative details. To counteract this, a Refine Net balances the original identity features with the aligned features. This ensures the model retains the fine-grained information necessary for distinguishing individuals while benefiting from the generalization properties of the aligned space.

3. Meta-Learning Strategy

To simulate domain shifts during training, the authors adopt a meta-learning approach. In each training iteration, source data is partitioned into "meta-train" and "meta-test" sets. The model is forced to learn features that remain effective across these simulated shifts, preparing it for the "shock" of a true unseen domain at deployment.

4. Test-Time Adaptive Updating

At deployment, MDA performs a rapid, "on-the-fly" update. Using a small number of unlabeled samples from the new environment, the model fine-tunes its distribution encoder to better align the local data distribution with the universal prior, further narrowing the domain gap.

Experimental Evaluation

The framework was evaluated on two challenging settings:

  • Large-scale DG ReID: Training on a combination of five datasets (CUHK02, CUHK03, Market-1501, DukeMTMC-ReID, and CUHK-SYSU) and testing on VIPeR, GRID, and i-LIDS.
  • Single-source DG ReID: Training on a single dataset (e.g., Market-1501) and testing on others to evaluate generalization from limited data.

Key Results

MDA significantly outperformed previous state-of-the-art (SOTA) methods:

  • Performance Gains: On the GRID dataset, MDA achieved a 12.8% improvement in Rank-1 accuracy over the previous best method (MetaBIN). Across all large-scale targets, the average Rank-1 increase was 5.2%.
  • Single-Source Robustness: Even with limited training data, MDA surpassed SOTA methods by up to 4.7% in Rank-1 accuracy.
  • Qualitative Analysis: t-SNE visualizations confirmed that MDA successfully clusters features from different domains into a shared region, whereas baseline models often produce isolated, domain-specific clusters.

Conclusion and Future Work

Meta Distribution Alignment provides a robust solution for DG ReID by leveraging a universal prior as a distributional anchor. While highly effective, the approach introduces some test-time latency due to adaptive updating, which may impact real-time constraints on edge devices. Future research could investigate more complex priors, such as Gaussian Mixtures, to better capture the diversity of human appearances while maintaining alignment benefits. Additionally, ethical considerations regarding privacy and tracking remain paramount in the deployment of ReID technologies.

Summary updated

Abstract

Domain Generalizable (DG) person ReID is a challenging task which trains a model on source domains yet generalizes well on target domains. Existing methods use source domains to learn domain-invariant features, and assume those features are also irrelevant with target domains. However, they do not consider the target domain information which is unavailable in the training phrase of DG. To address this issue, we propose a novel Meta Distribution Alignment (MDA) method to enable them to share similar distribution in a test-time-training fashion. Specifically, since high-dimensional features are difficult to constrain with a known simple distribution, we first introduce an intermediate latent space constrained to a known prior distribution. The source domain data is mapped to this latent space and then reconstructed back. A meta-learning strategy is introduced to facilitate generalization and support fast adaption. To reduce their discrepancy, we further propose a test-time adaptive updating strategy based on the latent space which efficiently adapts model to unseen domains with a few samples. Extensive experimental results show that our model outperforms the state-of-the-art methods by up to 5.2% R-1 on average on the large-scale and 4.7% R-1 on the single-source domain generalization ReID benchmark. Source code is publicly available at https://github.com/haoni0812/MDA.git.
BibTeX
@inproceedings{ni2022meta,
  author = {Hao Ni and Jingkuan Song and Xiaopeng Luo and Feng Zheng and Wen Li and Heng Tao Shen},
  title = {{Meta Distribution Alignment for Generalizable Person Re-Identification}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2022},
  doi = {10.1109/cvpr52688.2022.00252},
}
Figure