CORE-ReID V2: UDA for Person and Vehicle ReID

• The paper introduces CORE-ReID V2, which extends the original framework with a two-stage pipeline combining CycleGAN-based augmentation and mean-teacher adaptation.
• It employs dual attention modules—ECAB for local and SECAB for global feature refinement—to reduce pseudo-label ambiguity and enhance clustering stability.
• Empirical results on standard Person and Vehicle ReID benchmarks show significant improvements in mAP and Rank-1 accuracy over previous state-of-the-art methods.

Searching arXiv for the cited CORE-ReID papers to ground the article in the current record. CORE-ReID V2 is an enhanced Unsupervised Domain Adaptation framework for Object Re-Identification that builds on CORE-ReID and extends its scope from Person ReID to Vehicle ReID, with stated applicability to general Object ReID. It addresses the standard UDA setting in which a labeled source domain and an unlabeled target domain do not share identity classes, and it combines CycleGAN-based source-domain augmentation, a mean-teacher target adaptation stage, multi-view pseudo-labeling, and an attention-based Ensemble Fusion++ module intended to refine both local and global representations while reducing pseudo-label ambiguity (Nguyen et al., 6 Aug 2025).

1. Conceptual scope and relation to CORE-ReID

CORE-ReID V2 is positioned as a direct successor to CORE-ReID, retaining the original two-stage structure while revising several of its central components. The predecessor emphasized camera-aware pre-training, teacher–student learning, multi-view clustering, ECAB-based local refinement, and Bidirectional Mean Feature Normalization for Person ReID under UDA (Nguyen et al., 5 Aug 2025). CORE-ReID V2 preserves that overall logic but resolves three limitations identified for the earlier formulation: ECAB in the original framework enhanced only local features while leaving global features unrefined; the original framework supported only deeper backbones such as ResNet50, ResNet101, and ResNet152; and its design was limited to Person ReID, with camera-aware style transfer depending on known camera counts, which is impractical for some Vehicle ReID or generic-object settings (Nguyen et al., 6 Aug 2025).

The framework is therefore defined by two intertwined extensions. First, it generalizes the original Person ReID pipeline to Vehicle ReID and, in the authors’ formulation, to Object ReID more broadly. Second, it modifies the optimization and fusion machinery by introducing SECAB alongside ECAB, adopting greedy K-means++ initialization for clustering, and supporting lightweight backbones such as ResNet18 and ResNet34. A plausible implication is that CORE-ReID V2 should be understood less as a minor revision than as a re-scoped UDA ReID framework whose design goals include both cross-domain robustness and deployability.

A common misconception is that the “Object ReID” label implies extensive validation on arbitrary object categories. The reported evidence is narrower: the framework is experimentally evaluated on widely used UDA Person ReID and Vehicle ReID benchmarks, while evaluation on other object categories is explicitly left for future work (Nguyen et al., 6 Aug 2025).

2. Two-stage architecture and training logic

The overall workflow is split into source-domain pre-training and target-domain fine-tuning. In the pre-training stage, CycleGAN is used to synthesize style-aligned data that expands the labeled source domain. In the fine-tuning stage, a teacher–student architecture is initialized from the source-pretrained weights, and the student is optimized using pseudo-labels derived from multiple feature views; the teacher is updated by exponential moving average and is the only network used at inference time (Nguyen et al., 6 Aug 2025).

The target adaptation stage follows a mean-teacher update rule:

$$\rho_{\tau,i} = \eta \rho_{\tau,i-1} + (1-\eta)\rho_s,$$

with momentum $=0.999$. This EMA mechanism is used to stabilize representation learning under noisy pseudo-labels. The fine-tuning pipeline further decomposes features into one global branch and two local branches obtained from a top/bottom split, then applies Ensemble Fusion++ and BMFN before clustering and supervised optimization on pseudo-labels.

The paper also provides three explicit algorithmic components that frame training as more than a conventional augmentation-plus-clustering loop. Algorithm 1 performs Global Grayscale Transformation. Algorithm 2 performs Local Grayscale Transformation by converting a randomly selected rectangle to grayscale with probability $P_{local}$. Algorithm 3 replaces random K-means++ seeding with Greedy K-means++ seeding, deterministically selecting the next centroid from candidate sets to improve clustering stability and spread. Taken together, these components indicate that CORE-ReID V2 treats augmentation, clustering initialization, and fusion as co-dependent parts of the adaptation procedure rather than isolated heuristics.

3. Source-domain pre-training and synthetic domain alignment

The source-domain stage uses CycleGAN because it provides unpaired image-to-image translation that preserves structure while altering style. For Person ReID, the framework treats each camera view as a domain and performs camera-aware style transfer across camera pairs. For Vehicle ReID, where camera counts may be large or unknown, it uses domain-aware transfer between datasets rather than camera-pair transfer. Synthetic outputs preserve identity labels and are merged with real images for supervised source training (Nguyen et al., 6 Aug 2025).

For Person ReID, the reported camera-aware transfer coverage is exhaustive at the level of ordered camera pairs: Market-1501, with 6 cameras, requires $6 \times (6-1)=30$ models, while CUHK03, with 2 cameras, requires $2 \times (2-1)=2$ models. For Vehicle ReID, the reported domain-aware setup trains two models, VeRi-776$\rightarrow$VehicleID and VehicleID$\rightarrow$VeRi-776. In all cases, CycleGAN training resizes inputs to $286 \times 286$, crops to $256 \times 256$, uses Adam, sets the generator learning rate to $0.0002$ and the discriminator learning rate to $=0.999$0, runs 30 epochs at constant learning rate, and linearly decays over the next 20 epochs.

The source backbone removes the original fully connected layer and adds a BN layer and a new classification FC with $=0.999$1 classes. The channel dimension of the BN layer is backbone-dependent: $=0.999$2 or $=0.999$3. Source training combines cross-entropy classification and triplet verification, with preprocessing differentiated by modality: Person images are resized to $=0.999$4, Vehicle images to $=0.999$5, both with edge padding of 10 pixels followed by random crop, random horizontal flip with probability $=0.999$6, global grayscale with $=0.999$7, local grayscale with $=0.999$8, and RGB normalization to ImageNet mean and standard deviation. The source optimization schedule uses an initial learning rate of $=0.999$9, decays by $P_{local}$0 at epochs 40 and 70, runs for 350 epochs, includes 10 warmup epochs, and uses PK sampling with 32 identities and 4 images per identity, for a batch size of 128.

A notable design change relative to the original CORE-ReID is the explicit use of grayscale transformations as source-domain augmentation. This suggests that V2 targets color bias as a domain-shift factor in addition to camera or dataset style shift.

4. Target-domain adaptation, Ensemble Fusion++, and pseudo-labeling

Fine-tuning on the unlabeled target domain begins from the source-pretrained student and teacher. Features are extracted in three views: one global view and two local views obtained by splitting the feature representation into top and bottom components. These views are then processed by Ensemble Fusion++, which combines ECAB for local feature refinement and SECAB for global feature refinement, followed by BMFN to merge original and horizontally flipped views before clustering (Nguyen et al., 6 Aug 2025).

ECAB and SECAB have distinct roles. ECAB is applied to local features and produces channel attention from both max-pooling and average-pooling descriptors using a Shared Multilayer Perceptron with ReLU and a sigmoid gate. SECAB is a simplified, GPU-friendly variant tailored to global feature maps; it keeps the pooling-to-SMP-to-sigmoid pathway but outputs only the attention map, omitting the residual reweighting step. In the proposed configuration, ECAB refines local channels and SECAB refines global channels, after which element-wise multiplication fuses local attention outputs with refined global features to form the fused top and bottom representations used for clustering.

BMFN then merges original and horizontally flipped maps, normalizes them, and refines them before downstream use. In practical terms, BMFN is part of the mechanism by which multi-view clustering becomes less sensitive to left-right asymmetries and instability across augmentations. The clustering stage independently assigns pseudo-labels to the global, top, and bottom views using K-means with Euclidean distance, so each target sample receives three pseudo-labels. Greedy K-means++ initialization is used instead of random seeding, and the implementation details are unusually explicit: clustering uses a maximum of 100 iterations, a cluster minibatch size of 512, early stopping after 50 minibatches without inertia improvement, an empty-cluster reassignment ratio of 0.05, and seed pools of 1,500 global and 900 local features.

The target loss combines global classification, global triplet, and local softmax triplet terms, with weights $P_{local}$1, $P_{local}$2, $P_{local}$3, and $P_{local}$4. The target optimization schedule runs for 80 epochs with 400 iterations per epoch, uses Adam with weight decay $P_{local}$5, and fixes the learning rate at $P_{local}$6 for Person ReID and $P_{local}$7 for Vehicle ReID. The reduction/expansion rate for SECAB and ECAB is $P_{local}$8, and the number of hidden layers is $P_{local}$9.

5. Datasets, protocols, and reported results

The framework is evaluated on standard Person ReID and Vehicle ReID UDA benchmarks. For Person ReID, the reported datasets are Market-1501, CUHK03, and MSMT17. Market-1501 contains 32,668 images from 6 cameras, with 12,936 training images of 751 IDs and a test split of 3,368 query and 19,732 gallery images from 750 IDs. CUHK03 under the rigorous manual-box protocol contains 14,097 images of 1,467 persons, with 7,365 training images of 767 IDs and a test split of 1,400 query and 5,332 gallery images from 700 IDs. MSMT17 contains 126,441 boxes from 4,101 identities and 15 cameras, with 32,621 training images of 1,041 IDs and a test split of 11,659 query and 82,161 gallery images from 3,060 IDs. For Vehicle ReID, the reported datasets are VeRi-776, VehicleID, and VERI-Wild (Nguyen et al., 6 Aug 2025).

Evaluation uses mAP and Rank-$6 \times (6-1)=30$0 Accuracy, specifically Rank-1, Rank-5, and Rank-10. The paper states that no re-ranking is used and that single-query protocol is followed according to dataset standards.

Transfer or test setting	mAP	Rank-1
Market $6 \times (6-1)=30$1 CUHK	66.4	66.9
CUHK $6 \times (6-1)=30$2 Market	84.5	93.9
Market $6 \times (6-1)=30$3 MSMT	44.1	71.3
CUHK $6 \times (6-1)=30$4 MSMT	40.7	68.7
VehicleID $6 \times (6-1)=30$5 VeRi-776	49.50	80.15
VehicleID $6 \times (6-1)=30$6 VERI-Wild Test3000	40.2	76.6
VehicleID $6 \times (6-1)=30$7 VERI-Wild Test5000	34.9	70.2
VehicleID $6 \times (6-1)=30$8 VERI-Wild Test10000	27.8	62.1
VeRi-776 $6 \times (6-1)=30$9 VehicleID Test800	67.04	58.32
VeRi-776 $2 \times (2-1)=2$0 VehicleID Test1600	63.02	53.49
VeRi-776 $2 \times (2-1)=2$1 VehicleID Test2400	57.99	48.62

Within Person ReID UDA, the authors report improvements over PAOA+ of $2 \times (2-1)=2$2 mAP on Market$2 \times (2-1)=2$3CUHK and $2 \times (2-1)=2$4 mAP on CUHK$2 \times (2-1)=2$5Market. On MSMT17 transfers, the gains over CACL and PAOA+ are reported as $2 \times (2-1)=2$6 and $2 \times (2-1)=2$7 mAP on Market$2 \times (2-1)=2$8MSMT and CUHK$2 \times (2-1)=2$9MSMT, respectively. For Vehicle ReID, CORE-ReID V2 is reported to outperform MGR-GCL and MATNet+DMDU on VehicleID$\rightarrow$0VeRi-776, and to surpass DMDU on VeRi-776$\rightarrow$1VehicleID by $\rightarrow$2, $\rightarrow$3, and $\rightarrow$4 mAP on Test800, Test1600, and Test2400. The lightweight variant, CORE-ReID V2 Tiny based on ResNet18, reports Market$\rightarrow$5CUHK mAP $\rightarrow$6, CUHK$\rightarrow$7Market mAP $\rightarrow$8, Market$\rightarrow$9MSMT mAP $\rightarrow$0, VehicleID$\rightarrow$1VeRi-776 mAP $\rightarrow$2, and VehicleID$\rightarrow$3VERI-Wild Test3000 mAP $\rightarrow$4.

6. Ablations, efficiency profile, and limitations

The ablation results isolate the contributions of style transfer, SECAB, the ECAB/SECAB division of labor, clustering initialization, and backbone choice. Direct transfer without the full pipeline is substantially weaker, with reported examples of Market$\rightarrow$5CUHK mAP $\rightarrow$6 and VehicleID$\rightarrow$7VeRi-776 mAP $\rightarrow$8, supporting the claim that CycleGAN-based source augmentation materially improves downstream adaptation (Nguyen et al., 6 Aug 2025).

SECAB is specifically introduced to refine global features, and its removal degrades performance. The reported Market$\rightarrow$9CUHK mAP decreases from $286 \times 286$0 with SECAB to $286 \times 286$1 without SECAB, while VehicleID$286 \times 286$2VeRi-776 mAP decreases from $286 \times 286$3 to $286 \times 286$4. The attention-placement study further reports that applying ECAB to both local and global branches yields Market$286 \times 286$5CUHK mAP $286 \times 286$6, applying SECAB to both yields $286 \times 286$7, and the proposed combination—ECAB on local plus SECAB on global—yields the best value of $286 \times 286$8. Greedy K-means++ also shows measurable benefit over random seeding, improving VehicleID$286 \times 286$9VeRi-776 mAP from $256 \times 256$0 to $256 \times 256$1 and Market$256 \times 256$2CUHK mAP from $256 \times 256$3 to $256 \times 256$4. The number of clusters remains a sensitive hyperparameter, with reported peaks at 900 for Market$256 \times 256$5CUHK, 2000 for Market$256 \times 256$6MSMT, 500 for VehicleID$256 \times 256$7VeRi-776, and 700 for VeRi$256 \times 256$8VehicleID.

Backbone	Parameters	GFLOPs
ResNet-18	12.97M	1.18
ResNet-34	23.08M	2.35
ResNet-50	46.62M	5.10
ResNet-101	65.61M	7.58
ResNet-152	81.26M	10.61

The backbone analysis indicates that ResNet101 typically gives the best accuracy across Person and Vehicle transfers, while ResNet18 and ResNet34 provide substantially lower parameter counts and per-image GFLOPs. Because only the teacher model is used during inference, runtime complexity is reduced relative to training-time cost.

The limitations are stated directly. Performance remains dependent on pseudo-label quality and can degrade in highly noisy or complex scenarios. The framework is sensitive to the choice of $256 \times 256$9, requiring dataset-specific tuning. Camera-aware transfer still assumes knowledge of camera domains, and when such information is unavailable the method falls back to domain-aware transfer, which may be less fine-grained than camera-aware mappings. Finally, although the framework is described as applicable to general Object ReID, validation on additional datasets such as BV-Person, ENTIRe-ID, VRID-1, VRAI, Vehicle-Rear, and V2I-CARLA is identified as future work. A plausible implication is that the most firmly established contribution of CORE-ReID V2 lies in extending the original CORE-ReID formulation from Person ReID UDA to a broader multi-domain ReID setting while preserving a single training logic centered on synthetic alignment, mean-teacher adaptation, and attention-based fusion (Nguyen et al., 5 Aug 2025).