Aligned Divergent Pathways for Omni-Domain Generalized Person Re-Identification

TL;DR

This work defines Omni-Domain Generalization for Person Re-Identification (ODG-ReID) and presents Aligned Divergent Pathways (ADP), a backbone-agnostic, multi-branch framework designed to generalize across diverse domain configurations. ADP uses DyMAIN to induce maximal cross-domain style variation, PMoC to diversify learning-rate trajectories across branches, and DCML to align features in the inter-branch space, collectively yielding strong omni-domain performance. Empirically, ADP achieves state-of-the-art results on multi-source domain generalization and single-domain ReID benchmarks, while ablations confirm the contributions of each component and their synergistic effects. The approach offers practical benefits for real-world deployment where domain shifts are common, supporting robust ReID across unseen cameras and environments.

Abstract

Person Re-identification (Person ReID) has advanced significantly in fully supervised and domain generalized Person R e ID. However, methods developed for one task domain transfer poorly to the other. An ideal Person ReID method should be effective regardless of the number of domains involved in training or testing. Furthermore, given training data from the target domain, it should perform at least as well as state-of-the-art (SOTA) fully supervised Person ReID methods. We call this paradigm Omni-Domain Generalization Person ReID, referred to as ODG-ReID, and propose a way to achieve this by expanding compatible backbone architectures into multiple diverse pathways. Our method, Aligned Divergent Pathways (ADP), first converts a base architecture into a multi-branch structure by copying the tail of the original backbone. We design our module Dynamic Max-Deviance Adaptive Instance Normalization (DyMAIN) that encourages learning of generalized features that are robust to omni-domain directions and apply DyMAIN to the branches of ADP. Our proposed Phased Mixture-of-Cosines (PMoC) coordinates a mix of stable and turbulent learning rate schedules among branches for further diversified learning. Finally, we realign the feature space between branches with our proposed Dimensional Consistency Metric Loss (DCML). ADP outperforms the state-of-the-art (SOTA) results for multi-source domain generalization and supervised ReID within the same domain. Furthermore, our method demonstrates improvement on a wide range of single-source domain generalization benchmarks, achieving Omni-Domain Generalization over Person ReID tasks.

Figures (3)

• Figure 1: can be applied on Transformer and ResNet blocks. The internal structure of our module is illustrated in detail in the flow-diagram at the bottom of the figure.
• Figure 2: Comparing a standard cosine learning rate schedule (a) with our ensemble of Phased Mixture-of-Cosine (PMoC) schedulers. The mix of turbulent and stable schedules in PMoC encourages learning diverse features among different branches.
• Figure 3: A summary of our method, . Image inputs are processed into sequences of patches and passed through a backbone model, such as a Vision Transformer. The final blocks of the backbone are cloned to make a self-ensemble of several branches. Our own normalization module DyMAIN is added to all branch blocks, and each branch specialized with a different cosine learning rate schedule via Phased Mixture-of-Cosines. A dimensional consistency metric loss (DCML) is imposed to further align the features from different branches.