---
title: 'CORE-ReID: UDA for Person & Object ReID'
url: https://www.emergentmind.com/topics/core-reid
type: topic
---

# CORE-ReID: UDA for Person & Object ReID

Searching arXiv for CORE-ReID and closely related papers to ground the article in current literature.
CORE-ReID is a family of unsupervised domain adaptation (UDA) frameworks for person re-identification (ReID) that couples camera- or domain-aware source pre-training with pseudo-label-based target-domain fine-tuning under a teacher–student regime. In its 2025 formulation, CORE-ReID denotes “Comprehensive Optimization and Refinement through Ensemble Fusion in Domain Adaptation for Person Re-identification,” a method that uses CycleGAN-based source augmentation, multi-view feature extraction, multi-level clustering, learnable Ensemble Fusion, Efficient Channel Attention Block (ECAB), and Bidirectional Mean Feature Normalization (BMFN) to improve target-domain pseudo-label quality and downstream retrieval accuracy [2508.03064]. CORE-ReID V2 extends this design from person ReID to a broader object Re-identification setting, including Vehicle ReID, while adding SECAB, Greedy K-Means++, global and local grayscale augmentation, and support for lightweight backbones such as ResNet-18 and ResNet-34 [2508.04036]. In a broader conceptual sense, the emphasis on exploiting richer identity structure also resonates with cross-video identity-correlating pre-training for person ReID, exemplified by CION, which treats identity discovery as a progressive multi-level denoising problem and uses identity-guided self-distillation [2409.18569].

## 1. Problem setting and scope

CORE-ReID is designed for UDA in person ReID. The source domain \(S\) is a fully labeled ReID dataset, while the target domain \(T\) is unlabeled; the objective is to train on labeled source data and unlabeled target data so that the model generalizes to target identities, which are disjoint from source identities [2508.03064]. The central difficulty is domain shift, arising from changes in camera networks, illumination, background, resolution, clothing styles, and identity populations [2508.03064].

The original CORE-ReID paper positions the framework against several concrete deficiencies in prior UDA ReID pipelines. These include treating the source domain as a single distribution while ignoring intra-source camera style differences, generating noisy pseudo-labels through single-view clustering, and using suboptimal local–global feature fusion strategies that do not adequately exploit flipped-image invariance or fine-grained local cues [2508.03064]. The method therefore seeks to make pre-training camera-aware, stabilize pseudo-label learning with a teacher–student framework, derive multiple complementary pseudo-labels through multi-view clustering, and resolve ambiguity through a learnable Ensemble Fusion mechanism with ECAB and BMFN [2508.03064].

CORE-ReID V2 broadens the scope from person ReID to object Re-identification, explicitly covering Person ReID and Vehicle ReID and stating further applicability to Object ReID [2508.04036]. This extension preserves the two-stage UDA design—supervised source pre-training followed by unsupervised target fine-tuning—but augments the original framework with stronger clustering, stronger feature fusion, broader augmentation, and lighter backbones [2508.04036].

A plausible implication is that CORE-ReID should be understood less as a single fixed architecture than as a modular UDA ReID design pattern: source-domain appearance harmonization, teacher–student pseudo-label refinement, multi-view feature construction, and learned fusion under pseudo-label supervision.

## 2. Original CORE-ReID architecture

The original CORE-ReID pipeline has two stages. In source-domain pre-training, CycleGAN-based camera-aware style transfer is used to generate additional camera-style variants of source images, and a ResNet-101 based ReID backbone is trained on real plus style-transferred labeled images using identity classification loss and triplet loss [2508.03064]. In target-domain fine-tuning, both student and teacher are initialized from the pre-trained weights; global and local features are extracted from target images; Ensemble Fusion with ECAB and BMFN produces fusion features; mini-batch K-means is run on teacher global features and on fusion top and fusion bottom features; and the student is trained from the resulting pseudo-labels while the teacher is updated by exponential moving average (EMA) [2508.03064].

Architecturally, the student receives an unlabeled target image \(x_{T,i}\) and its horizontally flipped version \(x'_{T,i}\). Its last convolutional feature map is split horizontally into top and bottom halves, yielding local descriptors after GAP. The teacher uses the same backbone but does not split the last feature map; GAP yields a global descriptor \(T_{\text{global} \in \mathbb{R}^{C}}\) [2508.03064]. Ensemble Fusion then uses student local descriptors and the teacher global descriptor to produce fusion top and bottom features, and BMFN merges original and flipped descriptors by
\[
f_m = \frac{F_m + F'_m}{2 \,\|F_m + F'_m\|_2}.
\]
This is applied to global, local, and fusion features [2508.03064].

The clustering stage operates at three levels: teacher global features produce pseudo-labels \(y_{T,i,\text{global}}\), fusion top features produce pseudo-labels \(y_{T,i,\text{top}}\), and fusion bottom features produce pseudo-labels \(y_{T,i,\text{bottom}}\) [2508.03064]. These pseudo-labels are used asymmetrically: global pseudo-labels supervise global classification plus triplet losses, while local pseudo-labels supervise local soft-max triplet losses [2508.03064]. The teacher parameters are updated as
\[
\rho_{\tau,i} = \eta \rho_{\tau,i-1} + (1-\eta)\rho_{s,i},
\]
with \(\eta = 0.999\) [2508.03064].

At inference time, only the teacher is used, without the fusion module, to reduce cost [2508.03064]. This design makes the computational overhead primarily a training-time phenomenon rather than an inference-time one.

## 3. Source-domain pre-training and camera-aware synthesis

A defining component of CORE-ReID is camera-aware style transfer during source pre-training. For a source dataset with \(C\) cameras, the method trains CycleGAN models for all directed camera pairs, amounting to \(C(C-1)\) mappings. For Market-1501, where \(C=6\), this yields 30 models; for CUHK03, where \(C=2\), it yields 2 models [2508.03064]. Each real source image from camera \(c\) is translated into the styles of the other \(C-1\) cameras, and the synthetic images inherit the original identity label [2508.03064].

The paper states that CORE-ReID also merges the training and test sets of the source dataset into a “total training set,” using the combined data only as source rather than for evaluation [2508.03064]. The ResNet-101 backbone is then trained from ImageNet initialization on the augmented labeled source set with identity classification and triplet loss:
\[
L_{S,\text{ID} = \frac{1}{N_S}\sum_{i=1}^{N_S} L_{ce}(C_S(f(x_{s,i})), y_{s,i}),
\]
\[
L_{S,\text{triplet} = \sum_{i=1}^{N_S}
\max\big(0, \|f(x_{s,i}) - f(x^+_i)\|_2^2 - \|f(x_{s,i}) - f(x^-_i)\|_2^2 + m\big),
\]
\[
L_{S,\text{total} = L_{S,\text{ID} + \kappa L_{S,\text{triplet}.
\]
The reported setting uses \(\kappa = 1\) [2508.03064].

CycleGAN itself is presented with the standard adversarial, cycle-consistency, and identity losses:
\[
L_{\text{CycleGAN}(G,F,D_X,D_Y) =
L_{\text{GAN}(G, D_Y, X, Y) + L_{\text{GAN}(F, D_X, Y, X) + \lambda L_{cyc}(G,F),
\]
\[
L_{\text{identity}(G,F) = \mathbb{E}_{y\sim p_{data}(y)}[\|G(y) - y\|_1] + \mathbb{E}_{x\sim p_{data}(x)}[\|F(x) - x\|_1],
\]
\[
L_{\text{total}(G,F,D_X,D_Y) = L_{\text{CycleGAN}(G,F,D_X,D_Y) + L_{\text{identity}(G,F).
\]
These components are used to preserve identity-relevant content while varying camera style [2508.03064].

This pre-training strategy is explicitly motivated by a criticism of standard UDA ReID practice: source pre-training normally ignores camera-style heterogeneity inside the labeled source domain. CORE-ReID turns this heterogeneity into a supervised augmentation signal [2508.03064].

## 4. Multi-view pseudo-labeling, Ensemble Fusion, ECAB, and BMFN

The original CORE-ReID paper uses three kinds of target features: teacher global features, and top and bottom fusion features derived from local student features and teacher global features [2508.03064]. The key idea is that each target image receives three pseudo-labels, one per feature view, and that these labels are made more reliable by learning fusion features that integrate local discriminative structure into the teacher’s global representation [2508.03064].

The Ensemble Fusion module takes student local descriptors \(S_{\text{top}}, S_{\text{bottom}}\) and teacher global descriptor \(T_{\text{global}}\). For each local branch, ECAB produces a channel attention vector \(v_\ell\), which modulates the teacher global feature:
\[
\widetilde{T}^{\text{global}_{\text{top} = v_{\text{top} \odot T_{\text{global},
\qquad
\widetilde{T}^{\text{global}_{\text{bottom} = v_{\text{bottom} \odot T_{\text{global}.
\]
GAP and BN then produce fusion descriptors \(O_{\text{top}}, O_{\text{bottom}}\), and BMFN merges original and flipped versions:
\[
o_\ell = \frac{O_\ell + O'_\ell}{2\|O_\ell + O'_\ell\|_2},\quad \ell\in\{\text{top},\text{bottom}\}.
\]
These outputs are used for clustering and local triplet supervision [2508.03064].

ECAB is described as being inspired by CBAM but more expressive. For a feature \(S \in \mathbb{R}^{C \times H \times W}\), max pooling and average pooling are passed through a Shared MLP with \(h\) hidden layers; the outputs are summed and passed through a sigmoid to produce a channel attention map:
\[
v = (s_{\text{max} + s_{\text{avg}),\quad
S_a = \sigma(v).
\]
The paper specifies \(r=4\) and \(h=5\) in experiments [2508.03064].

BMFN averages original and flipped features and applies L2 normalization:
\[
f_m = \frac{F_m + F'_m}{2\|F_m + F'_m\|_2}.
\]
The intended effect is to reduce deviation caused by orientation reversal or background differences and to encourage identity-related invariance [2508.03064].

The target fine-tuning losses are:
\[
L^{\text{global}_{T,\text{ID} = \frac{1}{N_T}\sum_{i=1}^{N_T} L_{ce}(C_T(f^s_{\text{global}(x_{T,i})), y_{T,i,\text{global}),
\]
\[
L^{\text{global}_{T,\text{triplet} = \frac{1}{N_T}\sum_{i=1}^{N_T} \max(0, \|f^s_{\text{global}(x_{T,i}) - f^s_{\text{global}(x^+_i)\|_2^2 - \|f^s_{\text{global}(x_{T,i}) - f^s_{\text{global}(x^-_i)\|_2^2 + m),
\]
\[
L^{\ell}_{T,\text{triplet} =
- \frac{1}{N_T}\sum_{i=1}^{N_T} \log\frac{e^{-\|f^s_{\ell}(x_{T,i}) - f^s_{\ell}(x^+_i)\|_2}}{ e^{-\|f^s_{\ell}(x_{T,i}) - f^s_{\ell}(x^+_i)\|_2} + e^{-\|f^s_{\ell}(x_{T,i}) - f^s_{\ell}(x^-_i)\|_2}},
\]
with total loss
\[
L_{T,\text{total} =
\alpha L^{\text{global}_{T,\text{ID}
+ \beta L^{\text{global}_{T,\text{triplet}
+ \gamma L^{\text{top}_{T,\text{triplet}
+ \delta L^{\text{bottom}_{T,\text{triplet},
\]
where \(\alpha = 1\), \(\beta = 1\), \(\gamma = 0.5\), and \(\delta = 0.5\) [2508.03064].

A common misconception is that CORE-ReID is simply a standard mean-teacher UDA method with feature concatenation. The paper instead defines a more specific mechanism: global clustering on teacher features, local supervision from fusion-derived pseudo-labels, learned channel attention via ECAB, and bidirectional original/flip normalization via BMFN [2508.03064].

## 5. CORE-ReID V2 and generalization beyond person ReID

CORE-ReID V2 is presented as an enhanced framework that builds on the original CORE-ReID while extending it to Person ReID, Vehicle ReID, and, in principle, Object ReID [2508.04036]. It keeps the same high-level two-stage UDA paradigm but modifies both source pre-training and target-domain pseudo-labeling.

On the source side, V2 uses camera-aware style transfer for Person ReID and domain-aware style transfer for Vehicle ReID. For person datasets, each camera is treated as a style domain and \(C(C-1)\) CycleGANs are trained; for vehicle datasets, where camera IDs may be missing, a source–target domain-aware CycleGAN is used instead [2508.04036]. V2 also introduces two grayscale-based augmentations: global grayscale transformation with probability \(P_{\text{global}}\) and local grayscale transformation with probability \(P_{\text{local}}\), explicitly intended to weaken over-reliance on color [2508.04036].

On the target side, V2 replaces random K-Means initialization with Greedy K-Means++ initialization for more stable pseudo-labels [2508.04036]. The paper defines
\[
D(x, C) = \min_{c \in C}\|x - c\|_2^2,\quad
D(X, C) = \sum_{x\in X} D(x, C),
\]
and describes a greedy center-selection process that samples candidates with probability proportional to squared distance and selects the candidate that most reduces clustering cost [2508.04036]. This change is explicitly motivated as a means to reduce bad initializations and improve pseudo-label quality.

The feature-fusion mechanism is also extended. V2 introduces Ensemble Fusion++, which uses ECAB on local features and SECAB on global features [2508.04036]. Let \(S_{\text{top}}, S_{\text{bottom}}\) denote student local feature maps and \(T_{\text{global}}\) denote the teacher global feature map. V2 computes
\[
\tau_{\text{top} = \text{ECAB}(S_{\text{top}),
\qquad
\tau_{\text{bottom} = \text{ECAB}(S_{\text{bottom}),
\]
\[
\phi_{\text{global} = \text{SECAB}(T_{\text{global}),
\]
then combines local and global attention maps by
\[
Z'_{\text{top,global}
= \tau_{\text{top} \odot \phi_{\text{global},
\qquad
Z'_{\text{bottom,global}
= \tau_{\text{bottom} \odot \phi_{\text{global}.
\]
After GAP and BN, these become fused descriptors \(\omega_{\text{top}}, \omega_{\text{bottom}}\), which are then merged with flipped-image counterparts through BMFN [2508.04036].

SECAB is described as a simplified ECAB for global features: it uses max and average pooled descriptors and a shared MLP but outputs only the attention map, not a reweighted feature map [2508.04036]. This is intended to reduce computational cost relative to ECAB while refining global information before fusion.

V2 also supports ResNet-18, ResNet-34, ResNet-50, ResNet-101, and ResNet-152, whereas the original CORE-ReID concentrated on deeper backbones [2508.04036]. This explicitly shifts the framework toward scalability and efficiency.

## 6. Empirical results, implementation, and relation to adjacent ReID research

The original CORE-ReID paper evaluates on four UDA scenarios: Market \(\rightarrow\) CUHK, CUHK \(\rightarrow\) Market, Market \(\rightarrow\) MSMT17, and CUHK \(\rightarrow\) MSMT17 [2508.03064]. It reports, for example, that on Market \(\rightarrow\) CUHK the baseline obtains mAP 40.1 and R-1 67.3, while CORE-ReID reaches mAP 41.9 and R-1 69.5; on CUHK \(\rightarrow\) Market, the baseline obtains mAP 37.2 and R-1 65.5, while CORE-ReID reaches mAP 40.4 and R-1 67.3 [2508.03064]. On Market \(\rightarrow\) MSMT17, the baseline obtains mAP 55.2 and R-1 55.7, while CORE-ReID reaches mAP 62.9 and R-1 61.0; on CUHK \(\rightarrow\) MSMT17, the baseline obtains mAP 82.2 and R-1 92.0, while CORE-ReID reaches mAP 83.6 and R-1 93.6 [2508.03064].

The same paper reports ablations on the number of K-means clusters, ECAB, BMFN, and backbone depth. It states that 900 clusters work best for Market \(\rightarrow\) CUHK and CUHK \(\rightarrow\) Market, while 2500 clusters work best for Market \(\rightarrow\) MSMT17 and CUHK \(\rightarrow\) MSMT17 [2508.03064]. It also reports that using both ECAB and BMFN yields the best results across all tasks, and that ResNet-101 performs best among ResNet-50, ResNet-101, and ResNet-152 [2508.03064].

CORE-ReID V2 reports further gains. For Person ReID, it gives Market \(\rightarrow\) CUHK results of 66.4 mAP and 66.9 R-1, CUHK \(\rightarrow\) Market results of 84.5 mAP and 93.9 R-1, Market \(\rightarrow\) MSMT results of 44.1 mAP and 71.3 R-1, and CUHK \(\rightarrow\) MSMT results of 40.7 mAP and 68.7 R-1 [2508.04036]. For Vehicle ReID, it reports VehicleID \(\rightarrow\) VeRi-776 results of 49.50 mAP and 80.15 R-1, and VeRi-776 \(\rightarrow\) VehicleID Test800 results of 67.04 mAP and 58.32 R-1 [2508.04036]. The paper further states that Greedy K-Means++ gives consistent gains of about \(1\)–\(2\) mAP and about \(1\)–\(1.5\) R-1, that SECAB improves mAP by about \(1\)–\(1.5\) and R-1 by about \(0.5\)–\(1\), and that the full Ensemble Fusion++ configuration is superior to all-ECAB or all-SECAB variants [2508.04036].

Implementation details are also explicit. In the original CORE-ReID, CycleGAN uses Adam with learning rates 0.0002 for generators and 0.0001 for discriminators, batch size 8, and 50 epochs total; source pre-training uses initial learning rate 0.00035, 120 epochs with 10 warm-up epochs, and 32 identities \(\times\) 4 images per batch; target fine-tuning uses 80 epochs, 400 iterations per epoch, learning rate 0.00035, Adam with weight decay 0.0005, and dual Quadro RTX 8000 GPUs [2508.03064]. In V2, source pre-training runs for 350 epochs with 10-epoch warmup, the default backbone is ResNet-101, and fine-tuning uses 80 epochs \(\times\) 400 iterations, fixed learning rate 0.00035 for Person ReID and 0.00007 for Vehicle ReID, Adam with weight decay 0.0005, clustering mini-batch size 512, and max 100 K-means iterations [2508.04036].

Within the broader ReID literature, CORE-ReID sits at the intersection of style-transferred source augmentation, clustering-based UDA, and feature-fusion methods [2508.03064]. The original paper explicitly situates itself relative to SPGAN, PTGAN, PDA-Net, HHL, CamStyle, SSG, MMT, MEB-Net, UNRN, TJ-AIDL, MMFA, UCDA, and CASCL [2508.03064]. CORE-ReID’s distinctive claim is not merely that it uses any one of these ingredients, but that it integrates camera-aware data generation, robust teacher–student pseudo-labeling, multi-view and multi-level clustering, and attention-driven ensemble fusion into a single UDA system [2508.03064].

A related but distinct line of work is cross-video identity-correlating pre-training. CION addresses person ReID pre-training on large-scale internet videos by explicitly correlating identities across videos, defining noise jointly through intra-identity consistency and inter-identity discrimination, and learning denoised cross-video identity structures with identity-guided self-distillation [2409.18569]. This is not CORE-ReID, but it shows a parallel move in the literature toward richer identity structure before downstream adaptation or fine-tuning [2409.18569]. This suggests a broader research trend: UDA ReID performance is increasingly tied to how well methods model latent identity structure, whether by source-style harmonization, pseudo-label fusion, or cross-video identity correlation.

The main limitations reported for CORE-ReID concern dependence on the quality of CycleGAN style transfer and the computational cost of combining multiple CycleGAN models, teacher–student training, multi-view feature extraction, multi-level clustering, and fusion [2508.03064]. CORE-ReID V2 adds that generalization to other specialized datasets remains untested, that experiments are still restricted mainly to Person and Vehicle ReID, and that pseudo-label noise remains a core challenge despite Greedy K-Means++ and Ensemble Fusion++ [2508.04036]. These limitations indicate that future work is likely to focus on lighter source-style transfer, stronger pseudo-label denoising, domain-specific normalization, adaptive sampling, and additional regularizers such as contrastive learning or adversarial regularization [2508.03064] [2508.04036].

Source: https://www.emergentmind.com/topics/core-reid