FedBPrompt: Federated Domain Generalization Person Re-Identification via Body Distribution Aware Visual Prompts

Xin Xu; Weilong Li; Wei Liu; Wenke Huang; Zhixi Yu; Bin Yang; Xiaoying Liao; Kui Jiang

cs.CVMar 13, 2026

FedBPrompt: Federated Domain Generalization Person Re-Identification via Body Distribution Aware Visual Prompts

Authors: Xin Xu, Weilong Li, Wei Liu, Wenke Huang, Zhixi Yu, Bin Yang, Xiaoying Liao, Kui Jiang

Organizations: School of Computer Science and Technology, Wuhan University of Science and Technology, China · Nanyang Technological University, Singapore · NERC for Multimedia Software, School of Computer Science, Wuhan University, China · Changsha Bus Group, China ·

Abstract

Federated Domain Generalization for Person Re-Identification (FedDG-ReID) learns domain-invariant representations from decentralized data. While Vision Transformer (ViT) is widely adopted, its global attention often fails to distinguish pedestrians from high similarity backgrounds or diverse viewpoints -- a challenge amplified by cross-client distribution shifts in FedDG-ReID. To address this, we propose Federated Body Distribution Aware Visual Prompt (FedBPrompt), introducing learnable visual prompts to guide Transformer attention toward pedestrian-centric regions. FedBPrompt employs a Body Distribution Aware Visual Prompts Mechanism (BAPM) comprising: Holistic Full Body Prompts to suppress cross-client background noise, and Body Part Alignment Prompts to capture fine-grained details robust to pose and viewpoint variations. To mitigate high communication costs, we design a Prompt-based Fine-Tuning Strategy (PFTS) that freezes the ViT backbone and updates only lightweight prompts, significantly reducing communication overhead while maintaining adaptability. Extensive experiments demonstrate that BAPM effectively enhances feature discrimination and cross-domain generalization, while PFTS achieves notable performance gains within only a few aggregation rounds. Moreover, both BAPM and PFTS can be easily integrated into existing ViT-based FedDG-ReID frameworks, making FedBPrompt a flexible and effective solution for federated person re-identification. The code is available at https://github.com/leavlong/FedBPrompt.

Figures & tables

Figure 1: The Dual Challenges of Client Heterogeneity in Federated Person Re-ID. (a) In the FedDG-ReID setting, clients exhibit heterogeneous background and viewpoint distributions. (b) This data heterogeneity in turn leads to two critical failure modes: (Top) The model becomes easily distracted by dominant yet irrelevant backgrounds, causing false matches between different individuals. (Bottom) Diverse viewpoints severely misalign body parts of the same individual, which drastically reduces their feature similarity and results in mismatches.

Figure 2: An Overview of the Proposed FedBPrompt Framework. The framework consists of two main components. (Left) On each client, our FedBPrompt method injects learnable prompts to guide the model’s attention toward pedestrian features and away from the background. The core BAPM then learns structured, part-level representations to solve the feature misalignment problem. (Right) The framework supports two training strategies. In Full-Parameter training, the entire model (~86M) is communicated. In contrast, our proposed PFTS freezes the backbone and only communicates the lightweight prompts (~0.46M), drastically reducing communication costs while maintaining high performance.

			MS+C2+C3→M		MS+C2+M→C3		C2+C3+M→MS		Average
Model	Reference	Method	mAP	rank-1	mAP	rank-1	mAP	rank-1	mAP	rank-1
FedProx [ 15 ]	MLSys 2020	official	33.4	57.6	24.0	22.4	11.2	25.3	22.9	35.1
		+PFTS	37.9 ( $\uparrow$ 4.5)	61.6 ( $\uparrow$ 4.0)	28.5 ( $\uparrow$ 4.5)	27.4 ( $\uparrow$ 5.0)	15.7 ( $\uparrow$ 4.5)	36.9 ( $\uparrow$ 11.6)	27.4 ( $\uparrow$ 4.5)	42.0 ( $\uparrow$ 6.9)
		+BAPM	47.3 ( $\uparrow$ 13.9)	70.9 ( $\uparrow$ 13.3)	33.7 ( $\uparrow$ 9.7)	33.7 ( $\uparrow$ 11.3)	17.7 ( $\uparrow$ 6.5)	41.2 ( $\uparrow$ 15.9)	32.9 ( $\uparrow$ 10.0)	48.6 ( $\uparrow$ 13.5)
Mixstyle [ 46 ]	ICLR 2021	official	29.3	53.3	19.6	18.6	12.4	27.9	20.4	33.3
		+PFTS	34.5 ( $\uparrow$ 5.2)	58.7 ( $\uparrow$ 5.4)	23.9 ( $\uparrow$ 4.3)	22.9 ( $\uparrow$ 4.3)	17.5 ( $\uparrow$ 5.1)	40.4 ( $\uparrow$ 12.5)	25.3 ( $\uparrow$ 4.9)	40.7 ( $\uparrow$ 7.4)

Table 1: Comparison of different methods under protocol-1 . M: Market1501, C2: CUHK02, C3: CUHK03, MS: MSMT17. Average represents the average performance over three unseen domains.

	M+C2+C3		M+C2+C3		M+C2+C3
	$\rightarrow$ M		$\rightarrow$ C2		$\rightarrow$ C3
Methods	mAP	rank-1	mAP	rank-1	mAP	rank-1
FedProx	68.7	85.1	79.9	79.3	44.6	45.6
FedProx+BAPM	72.9	88.4	83.1	83.3	49.5	50.6
FedPav	66.5	83.2	78.5	78.5	42.3	42.9
FedPav+BAPM	74.3	88.5	84.2	84.1	52.2	53.0

Table 2: Comparison of different methods on Protocol-2 . Under this protocol, we focus on the model’s identification performance on the source domains, the most basic yet often overlooked capability by existing methods.

	Prompt Type		C2+C3+MS $\rightarrow$ M		C2+M+MS $\rightarrow$ C3		C2+C3+M $\rightarrow$ MS
Methods	Global	Local	mAP	rank-1	mAP	rank-1	mAP	rank-1
	$\times$	$\times$	46.3	69.6	33.7	33.4	20.0	43.7
	✓	$\times$	48.4	73.2	36.7	36.8	22.9	48.2
	$\times$	✓	47.7	71.8	37.5	37.9	22.7	48.5
SSCU	✓	✓	49.1	73.4	37.4	38.4	23.4	49.5
	$\times$	$\times$	39.6	64.5	32.5	33.0	13.7	30.1

Table 3: Ablation study on key components. We validate the effectiveness of different prompt types. “Global” and “Local” denote Holistic Full Body Prompts and Body Part Alignment Prompts, respectively. Their combination forms our full BAPM.

Figure 3: Attention maps on Market-1501 under severe cropping, misalignment, and occlusion. Unlike the scattered attention of the Baseline , our BAPM exhibits clear specialization: body part prompts ( $\mathbf{P}^{\mathrm{upper}}$ , etc. ) localize specific regions, while holistic prompts ( $\mathbf{P}^{\mathrm{full}}$ ) capture global features. This aggregation yields comprehensive and accurate BAPM attention.

Class Token		RISE
Method	Ins. AUC	Method	Ins. AUC
SSCU	0.6160	SSCU	0.6516
+VPs	0.7103	+VPs	0.7494
BAPM	0.7559	BAPM	0.7737

Table 4: Comparison of Attention Map Types (Ins. AUC). We evaluate the fidelity of different attention map generation methods using Insertion Area Under the Curve (Ins. AUC) metric.

Figure 4: T-SNE visualizations of feature distribution on CUHK02 (C2), CUHK03 (C3), and MSMT17 (MS).

Federated Domain Generalization

See all 10 papers

CO-EVO: Co-evolving Semantic Anchoring and Style Diversification for Federated DG-ReID
Apr 29, 2026Fengchun Zhang, Qiang Ma, Liuyu Xiang +3Cross-Domain Generalization Federated Domain Generalization
School of Information and Software Engineering, University of Electronic Science and Technology of China · 2QiYuan Lab · School of Artificial Intelligence, Beijing University of Posts and Telecommunications +1
FedARKS: Federated Aggregation via Robust and Discriminative Knowledge Selection and Integration for Person Re-identification
Mar 6, 2026Xin Xu, Binchang Ma, Zhixi Yu +1Federated Domain Generalization Person Re-Identification
School of Computer Science and Technology, Wuhan University of Science and Technology, Wuhan 430065 China
Multi-Grained Vision-Language Alignment for Domain Generalized Person Re-Identification
Mar 14, 2026Jiachen Li, Xiaojin Gong, Dongping ZhangPerson Re-Identification Vision-Language Alignment
College of Information Science and Electronic Engineering, Zhejiang University · College of Information Engineering, China Jiliang University

Apr 29, 2026J/K move · Enter open · S save

CO-EVO: Co-evolving Semantic Anchoring and Style Diversification for Federated DG-ReID

Fengchun Zhang, Qiang Ma, Liuyu Xiang, Jinshan Lai+2

School of Information and Software Engineering, University of Electronic Science and Technology of China · 2QiYuan Lab · School of Artificial Intelligence, Beijing University of Posts and Telecommunications+1

Cross-Domain Generalization Federated Domain Generalization Federated Learning Person Re-Identification

Open paper AskPDF arXiv

Federated domain generalization for person re-identification (FedDG-ReID) aims to collaboratively train a pedestrian retrieval model across multiple decentralized source domains such that it can generalize to unseen target environments without compromising raw data privacy. However, this task is significantly challenged by the inherent stylistic gaps across decentralized clients. Without global supervision, models easily succumb to shortcut learning where representations overfit to domain specific camera biases rather than universal identity features. We propose CO-EVO, a novel federated framework that resolves this semantic-style conflict through a co-evolutionary mechanism. On the semantic side, Camera-Invariant Semantic Anchoring (CSA) learns identity prompts with cross-camera consistency to establish purified and domain-agnostic anchors that filter out local imaging noise. On the visual side, Global Style Diversification (GSD), powered by a Global Camera-Style Bank (GCSB), synthesizes realistic perturbations to expand the visual boundaries of training data. The core of CO-EVO is its co-evolutionary loop where purified anchors act as gravitational centers to guide the image encoder toward robust anatomical attributes amidst diverse style variations. Extensive experiments demonstrate that CO-EVO achieves state-of-the-art (SOTA) performance, proving that the synergy between semantic purification and style expansion is essential for robust cross-domain generalization. Our code is available at: https://github.com/NanYiyuzurn/ACL-LGPS-2026.

Abstract

Federated Domain Generalization

Explore similar work

CO-EVO: Co-evolving Semantic Anchoring and Style Diversification for Federated DG-ReID

FedARKS: Federated Aggregation via Robust and Discriminative Knowledge Selection and Integration for Person Re-identification

Multi-Grained Vision-Language Alignment for Domain Generalized Person Re-Identification