Collaborative Style Training in FedDG-ReID

• Collaborative Style Training (CST) is a dual-branch strategy that rapidly adapts to new style variations while continuously leveraging validated positive styles.
• Its methodology combines new style adaptation with memory-based reinforcement, ensuring both branches contribute equally to model updates.
• The approach enhances federated domain generalization in person ReID, as evidenced by significant improvements in Rank-1 accuracy and mAP.

Searching arXiv for the target paper and related terminology. Collaborative Style Training (CST) is a training strategy introduced as the second pillar of the Style Screening and Continuous Utilization (SSCU) framework for federated domain generalization in person re-identification (FedDG-ReID) in "Positive Style Accumulation: A Style Screening and Continuous Utilization Framework for Federated DG-ReID" (Xu et al., 22 Jul 2025). Its stated objective is two-fold: to let each client model rapidly adapt to newly generated styles and to continuously exploit the positive styles accumulated in the Generalization Gain-guided Dynamic Style Memory (GGDSM). CST is motivated by the observation that not all styles contribute to generalization performance; styles beneficial or harmful to the model’s generalization performance are therefore defined as positive or negative styles.

1. Problem setting and conceptual role

FedDG-ReID aims to learn a global server model that can be effectively generalized to source and target domains through distributed source domain data. Existing methods mainly improve the diversity of samples through style transformation, which to some extent enhances the generalization performance of the model. The central premise behind CST is narrower: not all generated styles are equally useful, and some may be detrimental to cross-domain generalization (Xu et al., 22 Jul 2025).

Within SSCU, this premise produces two coupled requirements. First, beneficial styles must be screened rather than accepted indiscriminately. Second, once screened, they must be utilized continuously rather than only at the moment of generation. CST addresses the second requirement directly, while GGDSM provides the style-selection and accumulation mechanism. In that division of labor, CST functions as the optimization procedure that operationalizes positive-style reuse during local training.

A common misunderstanding is to treat CST as a generic style-augmentation routine. In the SSCU formulation, it is more specific: newly generated styles are only one branch of the training signal, and the other branch is driven by positive styles previously accumulated in memory. This distinction is foundational to the method’s claim that continuous utilization, rather than one-off augmentation, is highly beneficial for generalization performance.

2. Client-side architecture and two-branch organization

At each client, CST maintains two network copies, a style transformation module, and a style memory of category prototypes. The architecture is defined as follows (Xu et al., 22 Jul 2025).

Component	Role	Symbol
Client-local model	Local branch model	$f_{\mathrm{Client\text{-}L}^k}$
Client-global model	Shared branch model initialized from server	$f_{\mathrm{Client\text{-}G}^k}$
Style transformation module	Generates stylized images	$f_T^k$
Style memory	Stores category prototypes	$M_k=\{M_k^i\}_{i=1}^{P_k}$

During each local training round, CST runs two parallel branches. The New Style Adaptation Branch samples a mini-batch $B_k=\{x_k^i\}$ from the local dataset $D_k$, generates stylized images, and feeds them into the client-global model. The Positive Style Continuous Utilization Branch instead inputs the original images into both the client-local and client-global models and optimizes them with the style memory recognition loss.

This two-branch decomposition is the defining structural property of CST. Because both branches share the client-global model, newly generated styles and memorized positive styles act on overlapping parameters in the same iteration. This suggests that the method is designed to couple rapid adaptation to novel style variation with persistent reinforcement of styles that have already shown utility at the global level.

3. Optimization objectives

The stylization step in the new-style branch is

$\hat{x}_k^i = f_T^k(x_k^i).$

The branch then applies a standard classification objective—cross-entropy plus triplet loss—so that $f_{\mathrm{Client\text{-}G}^k}$ quickly absorbs the new style variations:

$$L_{\mathrm{NS}^k} = L_{\mathrm{CE}}\!\bigl(\mathrm{cls}_k(f_{\mathrm{Client\text{-}G}^k}(\hat{x}_k^i)), y_k^i\bigr) + L_{\mathrm{Tri}}\!\bigl(f_{\mathrm{Client\text{-}G}^k}(\hat{x}_k^i), y_k^i\bigr).$$

The positive-style branch is governed by the style memory recognition loss

$$\ell_{id}(f, x_k^i, M_k) = -\log \frac{ \exp\!\bigl(\|f(x_k^i)\|_2^\top M_k^i / \tau\bigr) }{ \sum_{n=1}^{P_k} \exp\!\bigl(\|f(x_k^i)\|_2^\top M_k^n / \tau\bigr) },$$

where $f_{\mathrm{Client\text{-}G}^k}$0 is a temperature, $f_{\mathrm{Client\text{-}G}^k}$1 denotes $f_{\mathrm{Client\text{-}G}^k}$2-normalization, and $f_{\mathrm{Client\text{-}G}^k}$3 is the prototype for identity $f_{\mathrm{Client\text{-}G}^k}$4. Using this loss, CST defines

$f_{\mathrm{Client\text{-}G}^k}$5

and then sums them:

$f_{\mathrm{Client\text{-}G}^k}$6

The collaborative style training loss is therefore

$f_{\mathrm{Client\text{-}G}^k}$7

No additional balancing hyperparameter is introduced, so newly generated styles and memorized positive styles each contribute equally to the update (Xu et al., 22 Jul 2025). In methodological terms, this is an unusually explicit design choice: the method does not interpolate the two objectives with a separate scalar weight, and the two branches always co-exist.

4. Federated-round execution and memory update mechanism

Within one federated round on client $f_{\mathrm{Client\text{-}G}^k}$8, CST follows a fixed loop. The client first downloads the latest global model, setting $f_{\mathrm{Client\text{-}G}^k}$9. The client-local model is initialized, if not already, so that it shares weights with $f_T^k$0. For each local minibatch $f_T^k$1, the client stylizes the batch, computes $f_T^k$2 on the stylized images, computes $f_T^k$3 on the original images, updates parameters of both $f_T^k$4 and $f_T^k$5 by descending $f_T^k$6, and collects style features $f_T^k$7 for memory update (Xu et al., 22 Jul 2025).

At the end of the round, the server aggregates all client-global models via FedAvg and evaluates its Rank-1 accuracy. Memory update is conditional on improvement of the global model relative to the previous round. If performance improves, each client updates its memory according to

$f_T^k$8

with momentum $f_T^k$9. Otherwise, the newly generated style features are discarded as negative and $M_k=\{M_k^i\}_{i=1}^{P_k}$0 remains unchanged.

The memory $M_k=\{M_k^i\}_{i=1}^{P_k}$1 enters the CST loop only through the style-memory recognition loss. Because that loss is applied at every iteration on both the client-local and client-global models, positive styles stored in memory are reinforced continuously. There is no separate scheduling beyond the evaluation trigger and the momentum coefficient $M_k=\{M_k^i\}_{i=1}^{P_k}$2; the two branches jointly shape the weight updates in every training iteration (Xu et al., 22 Jul 2025).

5. Empirical contribution of the two branches

The ablation study reported for Table 4 evaluates how each CST component contributes under the settings “MS + C2 + C3 $M_k=\{M_k^i\}_{i=1}^{P_k}$3 M” and “M + C2 + C3 $M_k=\{M_k^i\}_{i=1}^{P_k}$4 MS” (Xu et al., 22 Jul 2025).

Variant	MS + C2 + C3 $M_k=\{M_k^i\}_{i=1}^{P_k}$5 M	M + C2 + C3 $M_k=\{M_k^i\}_{i=1}^{P_k}$6 MS
Baseline (no CST)	mAP $M_k=\{M_k^i\}_{i=1}^{P_k}$7, rank-1 $M_k=\{M_k^i\}_{i=1}^{P_k}$8	mAP $M_k=\{M_k^i\}_{i=1}^{P_k}$9, rank-1 $B_k=\{x_k^i\}$0
+ New Style Adaptation only	mAP $B_k=\{x_k^i\}$1 $B_k=\{x_k^i\}$2, rank-1 $B_k=\{x_k^i\}$3 $B_k=\{x_k^i\}$4	mAP $B_k=\{x_k^i\}$5 $B_k=\{x_k^i\}$6, rank-1 $B_k=\{x_k^i\}$7 $B_k=\{x_k^i\}$8
+ Positive Style Continuous Utilization only	mAP $B_k=\{x_k^i\}$9 $D_k$0, rank-1 $D_k$1 $D_k$2	mAP $D_k$3 $D_k$4, rank-1 $D_k$5 $D_k$6
NSA + PSCU (full CST)	mAP $D_k$7 $D_k$8, rank-1 $D_k$9 $\hat{x}_k^i = f_T^k(x_k^i).$0	mAP $\hat{x}_k^i = f_T^k(x_k^i).$1 $\hat{x}_k^i = f_T^k(x_k^i).$2, rank-1 $\hat{x}_k^i = f_T^k(x_k^i).$3 $\hat{x}_k^i = f_T^k(x_k^i).$4

These results confirm two distinct effects stated by the paper: learning new styles rapidly and continuously reinforcing positive memory styles are both crucial. The full CST configuration exceeds either single-branch variant in both settings, which is consistent with the method’s two-branch design rather than a one-branch substitution.

The reported qualitative evidence is aligned with the ablation pattern. The screening strategy in Fig. 3a–b allows the global model to recover quickly from occasional performance drops by re-injecting positive styles, and t-SNE visualizations in Fig. 3c show progressively tighter, more domain-invariant feature clusters as training proceeds. The authors summarize the overall outcome as state-of-the-art generalization performance on both source and unseen target ReID domains (Xu et al., 22 Jul 2025).

6. Terminological ambiguity and broader interpretation

The acronym CST is not unique to federated ReID. In "The ASCCR Frame for Learning Essential Collaboration Skills" (Vance et al., 2018), the same acronym is used for a training program built around the ASCCR Frame—Attitude, Structure, Content, Communication, and Relationship. That usage concerns interdisciplinary collaboration skills for statisticians and data scientists, with components such as POWER, $\hat{x}_k^i = f_T^k(x_k^i).$5, and ADEPT, rather than style transformation, prototype memory, or federated optimization. The two meanings are unrelated despite the shared acronym.

Within the FedDG-ReID literature represented here, a second misconception is that style augmentation alone explains the gains. The CST formulation is narrower and more selective. Newly generated style features are retained only when the global model’s performance improves; otherwise they are discarded as negative. Positive styles are then reintroduced through the style-memory recognition loss at every iteration on both client models (Xu et al., 22 Jul 2025). This suggests that CST is best understood not as indiscriminate diversification, but as a coordinated mechanism for exploration of new style variation and exploitation of styles already validated by generalization gain.

A plausible implication is that CST’s significance lies in how it couples local training dynamics to a global evaluation signal. The local client does not merely generate more stylized samples; it trains under a memory whose contents are filtered by cross-round improvements in server-side Rank-1 accuracy. In that sense, CST is inseparable from GGDSM and from the broader SSCU objective of screening and continuous utilization rather than simple style expansion.