Low-Rank Expert Merging for Multi-Source Domain Adaptation in Person Re-Identification

TL;DR

The paper tackles multi-source domain adaptation for person re-identification without sharing source data. It introduces SAGE-reID, which learns source-specific low-rank adapters (LoRA) for each dataset in a source-free setting and then merges them with a lightweight, input-driven gating network that blends the adapters on a shared backbone. The two-stage approach—target-adapted LoRAs followed by gate-based fusion—achieves strong, parameter-efficient cross-domain transfer, outperforming state-of-the-art MSDA methods on standard benchmarks. This work significantly reduces memory and computation while maintaining high accuracy, making multi-source reID more practical in privacy-conscious deployments.

Abstract

Adapting person re-identification (reID) models to new target environments remains a challenging problem that is typically addressed using unsupervised domain adaptation (UDA) methods. Recent works show that when labeled data originates from several distinct sources (e.g., datasets and cameras), considering each source separately and applying multi-source domain adaptation (MSDA) typically yields higher accuracy and robustness compared to blending the sources and performing conventional UDA. However, state-of-the-art MSDA methods learn domain-specific backbone models or require access to source domain data during adaptation, resulting in significant growth in training parameters and computational cost. In this paper, a Source-free Adaptive Gated Experts (SAGE-reID) method is introduced for person reID. Our SAGE-reID is a cost-effective, source-free MSDA method that first trains individual source-specific low-rank adapters (LoRA) through source-free UDA. Next, a lightweight gating network is introduced and trained to dynamically assign optimal merging weights for fusion of LoRA experts, enabling effective cross-domain knowledge transfer. While the number of backbone parameters remains constant across source domains, LoRA experts scale linearly but remain negligible in size (<= 2% of the backbone), reducing both the memory consumption and risk of overfitting. Extensive experiments conducted on three challenging benchmarks: Market-1501, DukeMTMC-reID, and MSMT17 indicate that SAGE-reID outperforms state-of-the-art methods while being computationally efficient.

Figures (4)

• Figure 1: Illustration of our SAGE‑reID method. Stage 1: Single-source UDA with LORA: each pre-trained source-specific model is augmented with low-rank adapters, and is finetuned on the target domain. Stage 2: LoRA fusion: a lightweight gating network predicts a linear combination of weights for each adapter and merges them on-the-fly in a single forward pass.
• Figure 2: Illustration of the proposed low-rank expert merging stage (Stage 2) in our MSDA framework. Each source-specific LoRA expert, trained independently in Stage 1, provides a low-rank residual update. During training, a lightweight gating network dynamically predicts linear combination of coefficients for each expert at every transformer layer based on the input features. These coefficients are then used to combine the residuals, enabling fusion of multi-source knowledge through a single forward pass. The backbone and LoRA weights remain frozen, making this stage computationally light while enhancing generalization to the target domain.
• Figure 3: Average gating coefficient $\bar{g}_{i,m}$ per target domain.
• Figure 4: Overall gating-correlation visualization across layers.