Generalizable Person Re-identification via Balancing Alignment and Uniformity

Abstract

Domain generalizable person re-identification (DG re-ID) aims to learn discriminative representations that are robust to distributional shifts. While data augmentation is a straightforward solution to improve generalization, certain augmentations exhibit a polarized effect in this task, enhancing in-distribution performance while deteriorating out-of-distribution performance. In this paper, we investigate this phenomenon and reveal that it leads to sparse representation spaces with reduced uniformity. To address this issue, we propose a novel framework, Balancing Alignment and Uniformity (BAU), which effectively mitigates this effect by maintaining a balance between alignment and uniformity. Specifically, BAU incorporates alignment and uniformity losses applied to both original and augmented images and integrates a weighting strategy to assess the reliability of augmented samples, further improving the alignment loss. Additionally, we introduce a domain-specific uniformity loss that promotes uniformity within each source domain, thereby enhancing the learning of domain-invariant features. Extensive experimental results demonstrate that BAU effectively exploits the advantages of data augmentation, which previous studies could not fully utilize, and achieves state-of-the-art performance without requiring complex training procedures.

Motivation

Data augmentation is a straightforward solution to enhance generalization capability by simulating diverse data variations during training. Due to its simplicity and effectiveness, numerous efforts have been made to adopt this approach for various DG tasks. However, in the context of DG re-ID, some data augmentations have been observed to exhibit a polarized effect — improving performance in the source domain while potentially degrading it in the target domain. We analyze this phenomenon through alignment and uniformity, revealing that it leads to sparse representation spaces with reduced uniformity.

Fig. Analysis on the polarized effect of data augmentations on in-distribution (ID) and out-of-distribution (OOD). (a) mAP (%) on Market-1501 of models trained on the same dataset (ID) and MS+CS+C3 (OOD) with varying augmentation probabilities. (b) Alignment \((\mathcal{L}_{\text{align}})\) and uniformity \((\mathcal{L}_{\text{uniform}})\) of OOD scenarios (MS+CS+C3 \(\rightarrow\) M). Counterintuitively, augmentations lead to more alignment but less uniformity, indicating that the model fails to sufficiently preserve the diverse information from the data distribution. (c) Uniformity \((- \mathcal{L}_{\text{uniform}})\) vs. augmentation probability for the source and target datasets in MS+CS+C3 \(\rightarrow\) M. Higher probabilities result in less uniformity, especially under distributional shifts, indicating an insufficiency for representing OOD data.

Balancing Alignment and Uniformity (BAU)

Based on our analysis, we propose a simple yet effective framework, Balancing Alignment and Uniformity (BAU), which mitigates the polarized effect of data augmentations by maintaining a balance between alignment and uniformity. Specifically, it regularizes the representation space by applying alignment and uniformity losses to both original and augmented images. Additionally, we introduce a weighting strategy that considers the reliability of augmented samples to improve the alignment loss. We further propose a domain-specific uniformity loss to promote uniformity within each source domain, enhancing the learning of domain-invariant features. Consequently, BAU effectively exploits the advantages of data augmentation, which previous studies could not fully utilize, and achieves state-of-the-art performance on various benchmarks.

Fig. Overview of the proposed framework. In (b) and (c), each color represents a different identity and domain, respectively. (a) With original and augmented images, we apply alignment and uniformity losses to balance feature discriminability and generalization capability. We further introduce a domain-specific uniformity loss to mitigate domain bias. (b) \(\mathcal{L}_{\mathrm{align}}\) pulls positive features closer, while \(\mathcal{L}_{\mathrm{uniform}}\) pushes all features apart to maintain diversity. (c) \(\mathcal{L}_{\text{domain-uniform}}\) uniformly distributes each domain's features and prototypes, reducing domain bias and thus enhancing generalization.

Expreimental Results

Table. Comparison with state-of-the-art methods on Protocol-1.

Table. Comparison with state-of-the-art methods on Protocol-2 and Protocol-3.

Fig. Analysis of alignment and uniformity. (a) Alignment \((\mathcal{L}_{\mathrm{align}})\) and uniformity \((\mathcal{L}_{\mathrm{uniform}})\) on Market-1501 when MS+CS+C3 \(\rightarrow\) M under Protocol-3 with varying augmentation probabilities. (b) T-SNE visualization with and without the domain-specific uniformity loss \(\mathcal{L}_{\mathrm{domain}}\). The values in parentheses in each legend label indicate the uniformity of the corresponding domain.

BibTeX


      @inproceedings{
        cho2024generalizable,
        title={Generalizable Person Re-identification via Balancing Alignment and Uniformity},
        author={Yoonki Cho and Jaeyoon Kim and Woo Jae Kim and Junsik Jung and Sung-Eui Yoon},
        booktitle={Advances in Neural Information Processing Systems},
        year={2024},
      }