Beyond Seen Bounds: Class-Centric Polarization for Single-Domain Generalized Deep Metric Learning

TL;DR

This work tackles Single-Domain Generalized Deep Metric Learning (SDG-DML), where models must generalize to both unseen categories and unseen domains. It proposes CenterPolar, a two-phase framework with Class-Centric Centrifugal Expansion (C$^3$E) to broaden domain coverage and Class-Centric Centripetal Constraint (C$^4$) to reinforce intra-class compactness and inter-class separation, leveraging geodesic distances on the hypersphere and class centroids. The approach yields theoretical guarantees on bounded expansion and stable optimization, and empirical results on five benchmarks (CUB-Ext, Cars196-Ext, DomainNet, PACS, Office-Home) demonstrate superior generalization and faster training than the state-of-the-art SEE method, with notable gains in MAP@R and RP. CenterPolar thus offers a principled, efficient mechanism to learn domain-invariant, category-discriminative metrics, with potential extensions to multi-modal and federated SDG-DML scenarios.

Abstract

Single-domain generalized deep metric learning (SDG-DML) faces the dual challenge of both category and domain shifts during testing, limiting real-world applications. Therefore, aiming to learn better generalization ability on both unseen categories and domains is a realistic goal for the SDG-DML task. To deliver the aspiration, existing SDG-DML methods employ the domain expansion-equalization strategy to expand the source data and generate out-of-distribution samples. However, these methods rely on proxy-based expansion, which tends to generate samples clustered near class proxies, failing to simulate the broad and distant domain shifts encountered in practice. To alleviate the problem, we propose CenterPolar, a novel SDG-DML framework that dynamically expands and constrains domain distributions to learn a generalizable DML model for wider target domain distributions. Specifically, \textbf{CenterPolar} contains two collaborative class-centric polarization phases: (1) Class-Centric Centrifugal Expansion ($C^3E$) and (2) Class-Centric Centripetal Constraint ($C^4$). In the first phase, $C^3E$ drives the source domain distribution by shifting the source data away from class centroids using centrifugal expansion to generalize to more unseen domains. In the second phase, to consolidate domain-invariant class information for the generalization ability to unseen categories, $C^4$ pulls all seen and unseen samples toward their class centroids while enforcing inter-class separation via centripetal constraint. Extensive experimental results on widely used CUB-200-2011 Ext., Cars196 Ext., DomainNet, PACS, and Office-Home datasets demonstrate the superiority and effectiveness of our CenterPolar over existing state-of-the-art methods. The code will be released after acceptance.

Figures (9)

• Figure 1: Comparison of different solutions in SDG-DML. (a) Solutions: A comparison of the existing class proxy-based expansion method (i.e., the SEE method) with our class center-based expansion method (i.e., our CenterPolar method). (b) Training: Corresponding to the solutions, we visualized the expansion and constraint processes during training using t-SNE. (c) Testing: The test results obviously show that our method can improve the generalization of the learned metric model. (Best viewed in color).
• Figure 2: An overview of the CenterPolar framework, which consists of two phases: C$^3$E (i.e., Class-Centric Centrifugal Expansion) and C$^4$ (i.e., Class-Centric Centripetal Constraint). C$^3$E expands the source domain by pushing samples away from class centroids while preserving semantic information; C$^4$ pulls all samples toward class centroids and enforces intra-class compactness and clearer inter-class separation.
• Figure 3: Ablation study results of hyperparameters ($m$ and $\lambda$). The left two parts on CUB-200-2011 Ext. (R $\rightarrow$ W): Results on different values of $\lambda$ with $m = 1$ and different values of $m$ with $\lambda = 0.75$, respectively. The right two parts show the impact of hyperparameter $\lambda$ when $m = 1$ on PACS and Office-Home datasets.
• Figure 4: Ablation study of different embedding dimensions, reporting R@1 and MAP@R. The left two parts present results on CUB-200-2011 Ext. (R $\rightarrow$ W); the right two parts show results on Cars196 Ext. (R $\rightarrow$ O).
• Figure 5: t-SNE visualization of the feature space learned from Baseline and CenterPolar on CUB-200-2011 Ext. (R $\rightarrow$ W), Cars196 Ext. (R $\rightarrow$ O), and DomainNet (R $\rightarrow$ P) datasets, respectively.