Shared Fixed Identity Geometry for Federated Person Re-Identification

Person re-identification means finding the same pedestrian again across different cameras. It is a natural fit for federated learning, which trains one model across many sites without moving raw images: each site trains locally, and a server averages the resulting weights. The catch is that every site sees different people. Person 17 at one camera network has nothing to do with person 17 at another. The standard workaround, federated partial averaging, shares only the feature extractor and gives each site its own private identity classifier that is never averaged. This avoids label clashes, but the private classifiers are free to settle into arbitrary, site-specific arrangements. The backbones they train then pull in incompatible directions when averaged.

We propose to remove that freedom. The server hands every site a fixed, never-trained set of classifier vectors, drawn from one global arrangement of nearly equiangular unit directions, and assigns each local identity its own vector. Sites still never share identities or images. They do, however, train toward the same target geometry in the same coordinates, and that geometry spreads identities across the whole embedding space. We expect this to reduce two failures that generic federated learning research has documented but nobody has measured in re-identification: classifier bias that corrupts the shared backbone, and "dimensional collapse," where features shrink into a low-dimensional subspace. We further expect the reduction to show up as better retrieval on cameras never seen during training.

Core Claim

Two results from generic federated classification motivate this work. No Fear of Classifier Biases (FedETF) shows that skewed local classifiers misalign the feature extractor. It fixes this by replacing the learned classifier with a fixed simplex equiangular tight frame, a set of class vectors with equal, maximal pairwise separation. That change raises global accuracy on highly skewed CIFAR-10 from 44.48% to 56.08%. Towards Understanding and Mitigating Dimensional Collapse (FedDecorr) proves that heterogeneous local data drives feature covariance toward low rank. Penalizing feature correlation gains about 8% over plain federated averaging on CIFAR-10.

Both papers assume a shared, closed label space. Neither applies directly to re-identification, where label spaces are disjoint and open-set, and where the test task is retrieval of people who were never seen in training. FedReIDBench/FedPav established disjoint labels as the defining obstacle and settled on unshared local heads. No one has checked whether those heads cause collapse.

Our claim is that a single globally coordinated fixed frame can serve disjoint identity sets. The frame is a large spherical code with low mutual coherence, and each client receives a disjoint slice of it. Such a frame preserves the benefit of FedETF in the open-set regime. The benefit should be measurable as higher effective rank of embeddings on unseen domains and as higher cross-domain mAP (mean average precision over the retrieval ranking).

The mechanism is simple. The server holds unit vectors $m_1,\dots,m_N$ in $\mathbb{R}^d$, chosen to minimize the coherence $\mu=\max_{i\neq j}|m_i^\top m_j|$. Here $N$ is the total identity count across clients, and it may exceed the embedding dimension $d$. In that case the Welch bound $\mu \ge \sqrt{(N-d)/(d(N-1))}$ says perfect equiangularity is impossible but coherence near $1/\sqrt{d}$ is attainable. Client $k$, with local identities $\mathcal{Y}_k$, receives the disjoint subset $\{m_j\}_{j\in\mathcal{Y}_k}$, assigned at random so that every client's slice spans the whole space rather than one subspace. It trains only the backbone $f_\theta$ with

$$\mathcal{L}_k=-\log\frac{\exp(s\,m_{y}^\top \hat f_\theta(x))}{\sum_{j\in\mathcal{C}}\exp(s\,m_j^\top \hat f_\theta(x))}+\mathcal{L}_{tri},$$

where $\hat f$ is the L2-normalized embedding, $s$ is a temperature, $y$ is the local identity of image $x$, and $\mathcal{L}_{tri}$ is the usual triplet loss. The negative set $\mathcal{C}$ is either the client's own slice or the full frame; the second option treats other clients' identity vectors as known-different people. Clients reveal only their identity counts, and only the backbone (or prompts) is averaged.

Research Program

The work proceeds in three steps.

Diagnosis. We will measure whether dimensional collapse occurs in federated re-identification at all. We will record the singular value spectrum and effective rank of embedding covariance for three models: centralized training, FedPav with learned local heads, and isolated local models. Spectra will be taken on each client's data and on held-out target domains, across rounds and local epochs. If FedPav embeddings are not measurably lower-rank than centralized ones, the premise weakens. In that case the paper becomes a negative-result study, and the next two steps would be de-emphasized.

Method and ablations. We will implement the coordinated fixed frame. Ablations will separate three factors: fixing the head at all, coordinating the frame globally versus each client drawing its own random vectors, and local versus full-frame negatives.

Regimes and composition. We will test two backbone regimes. The first trains a full ResNet-50, as in FedPav, DFH and DACS. The second freezes a ViT and trains only prompts, as in FedBPrompt. A frozen foundation backbone may already resist collapse, and we want to find out whether it does.

We will also check composition. DFH shares channel-wise feature statistics and interpolates them into hallucinated domains. DACS trains style perturbations constrained by a local-expert/global-model entropy test. Both act on the input or feature style and leave the classifier head as in FedPav. Our change targets only the head geometry, so it should stack with them. We will add it to DACS to test whether the gains are complementary rather than overlapping.

FedCAPR matters here only as a constraint. Its unsupervised clustering produces a new label set every round, which would require reassigning frame slices each round. We treat that as an extension, not the core.

Baselines

The naive reference is FedPav with learned, unshared local heads and sample-count averaging, run under the federated domain generalization protocol used by DACS. The protocol trains on CUHK02, CUHK03 and MSMT17 as three clients and tests on Market-1501. On that protocol DACS reports 25.4% mAP for FedPav and 30.1% for the FedReID baseline. We will also include isolated local training and plain federated averaging. DFH reports 61.3% mAP for federated averaging on CUHK02 as target, against 71.7% for DFH.

The strongest existing competitor on the main protocol is DACS, which DACS reports at 36.3% mAP on Market-1501. We will reproduce it with the authors' settings before comparing. A reproduced number more than about one point below 36.3% would be flagged, and all methods would then be compared against our reproduction.

In the frozen-ViT regime the competitor is FedBPrompt. For the Market-1501, CUHK02 and CUHK03 to MSMT17 split, the file reports FedBPrompt only as a +3.4% mAP and +5.8% Rank-1 improvement over its own baseline, with no absolute value. We will therefore report our own reproduced FedBPrompt number.

FedDecorr is the most direct generic competitor for the collapse question. We found no reported result on Lacuna for FedDecorr on person re-identification, so we will run it on top of FedPav with its correlation penalty weight tuned.

Every method gets the same backbone, clients, rounds, local epochs and augmentation, and an equal hyperparameter search of about twelve configurations. Tuning uses a validation split carved from the source clients, never the target. Beating the strongest baseline means improving target mAP by more than the seed-to-seed standard deviation over three seeds, both alone and when added to DACS.

The key ablation is fixed but uncoordinated random heads, where each client draws its own vectors. If this matches the coordinated frame, the contribution is "fix the head," not "share the geometry," and we will say so.

Why not just…?

Why not just add FedDecorr's decorrelation penalty to FedPav? This is the most serious threat. It is cheap, needs no server coordination, and directly targets collapse. It spreads features across dimensions but leaves the learned local heads free to bias the backbone. It also says nothing about where different clients' identities should sit relative to each other. The head-to-head comparison with a matched tuning budget, plus the stacked combination, will settle it. If FedDecorr alone matches our method on target mAP and effective rank, we will report the fixed frame as offering no advantage.

Why not just use metric losses only and drop classifiers entirely? Triplet or contrastive training with no head sidesteps disjoint labels. However, re-identification practice combines cross-entropy with triplet loss because the pairwise loss alone learns weaker embeddings. FedPav's design choice also implies its authors found a classification signal worth keeping. We will run a triplet-only federated baseline to measure the gap directly rather than assume it.

Why not just freeze a foundation backbone and tune prompts, as FedBPrompt does? If the backbone barely moves, averaging cannot collapse it, and the head geometry might not matter. This is exactly why the frozen-ViT regime is included. A null result there, alongside a gain in the ResNet regime, would still be informative: it would tell practitioners that coordinated heads matter only when the backbone is trained.

Why not just give each client its own fixed random head? In high dimensions, independently drawn vectors are already nearly orthogonal, so coordination may add little. The uncoordinated ablation tests this. We expect coordination to matter mainly when the total identity count greatly exceeds $d$; MSMT17 alone has 4,101 identities against a 768- or 2,048-dimensional embedding. We will vary $d$ to check this.

Proposed Experiments

The main table covers the three-source to Market-1501 protocol from DACS and the rotation to CUHK02 as target used by DFH, in both backbone regimes. It reports mAP and Rank-1 over three seeds.

The diagnostic figure plots target-domain effective rank against target mAP for every method. It asks whether higher rank predicts better transfer across methods, which is the mechanism our claim rests on.

A second setting uses the federated-by-camera split of CUHK03-NP from FedReIDBench, where naive training reaches 11.21% Rank-1. Each camera there has few identities, so classifier bias should be worst and our method's gain should be largest.

The decisive checks are three. First, FedPav embeddings must show measurably lower effective rank than centralized ones. Second, the coordinated frame must beat FedDecorr-on-FedPav on target mAP. Third, it must add gains on top of DACS rather than merely matching it. If the first check fails, the method loses its motivation. If the second fails, the simpler regularizer wins.