SCM-ReID: GAN-based Domain Adaptation

• SCM-ReID is a framework that combines GAN-based image translation and IQA reweighting to address appearance variations in cross-domain person re-identification.
• IQAGA utilizes StarGAN for style adaptation and a ResNet-50 backbone with IQA-weighted loss to achieve 32–36% mAP improvements on benchmarks like Market→Duke.
• DAPRH builds on IQAGA by integrating domain-invariant mapping, Vision Transformer-based holistic features, and pseudo-label refinement to boost mAP to over 70% on Market→Duke.

Person re-identification (ReID) under domain shift is a principal challenge for intelligent surveillance, primarily due to appearance variations and domain discrepancies between cameras or datasets. IQAGA (Image Quality–driven GAN Augmentation) and DAPRH (GAN Augmentation + Pseudo-Label Refinement + Holistic features) are two advanced unsupervised domain adaptation (UDA) methods that integrate generative augmentation, domain adaptation, and discriminative learning to improve cross-domain ReID performance. Both methods leverage StarGAN-based image translation, but they differ in their domain alignment and target-domain learning strategies, achieving substantial gains in mAP and Rank-1 accuracy compared to prior UDA approaches (Pham et al., 4 Jan 2026).

1. Architectural Overview and Workflow

1.1 IQAGA

IQAGA adopts a two-stage workflow:

• Stage I (Image-Level Adaptation): StarGAN is trained to translate source images $x_s$ into $C_t$ target camera styles $c$ by optimizing adversarial loss ($L_{adv}$), domain classification ($L_{cls}$), cycle reconstruction ($L_{rec}$), identity mapping ($L_{idt}$), and identity-preserving color loss ($L_{pid}$). For each $x_s$, $C_t$ style-transferred images are generated and combined with the original source set to form $D_{sync}$.
• Stage II (Supervised Training on Source): A ResNet-50 backbone extracts 2048-dimensional features. Supervised learning is performed using cross-entropy ($L_{ce}$) and triplet loss ($L_{tri}$), with sample-level reweighting based on normalized image quality assessment (IQA) scores $z_i$. The IQA-weighted loss ($L_i$) increases the contribution of high-quality samples.

1.2 DAPRH

DAPRH extends IQAGA with additional target-side adaptation modules:

• Stage I (Source Training): StarGAN-based augmentation is used, but only a fraction $N:M$ of real:GAN images per batch (e.g., 4:1) are included to reduce GAN-induced noise. A domain-invariant mapping (DIM) module employs adversarial training between a domain classifier $D_{net}$ and the feature extractor $fe$ to learn domain-confused features.
• Stage II (Target Unsupervised Training): Unlabeled target features are extracted and clustered (DBSCAN) to assign pseudo-labels $y_i$. Enhanced queries $v_i$ are generated through a Vision Transformer (ViT), integrating global (CLS token) and top-K local features. Pseudo-label refinement (CRL) updates labels with soft cluster probabilities and silhouette scores. Further, a teacher-student framework with EMA parameter transfer, camera-aware proxies (CAP), and contrastive objectives are employed for robust target learning.

2. Loss Functions and Optimization

2.1 IQAGA Losses

• StarGAN Generator Loss:

$$L_G = L_{adv}(G) + \lambda_{cls} L_{cls}(G) + \lambda_{rec} L_{rec} + \lambda_{idt} L_{idt} + \lambda_{pid} L_{pid}$$

• Discriminator Loss:

$L_D = L_{adv}(D) + \lambda_{cls} L_{cls}(D)$

• Source Training Loss:

For each sample, $$z_i = \text{clamp}\left(\frac{f_i - \mu_f}{\sigma_f / h}, -1, 1\right)$$ with $h=0.33$; $L_i = (1 + A_z z_i) \cdot (L_{ce,i} + L_{tri,i})$ Total loss: $L_{source} = \sum_i L_i$

2.2 DAPRH Losses

• Domain-Invariant Mapping:

$$L_D = \mathbb{E}_{s \in source}[(D_{net}(f_s) - 1)^2] + \mathbb{E}_{t \in target}[D_{net}(f_t)^2]$$

$$L_{DIM} = \mathbb{E}_{s}[(D_{net}(f_s) - 0.5)^2] + \mathbb{E}_{t}[(D_{net}(f_t) - 0.5)^2]$$

• Target Unsupervised Loss:
  • $L_{NCE}$: Instance-to-cluster contrastive loss
  • $L_{CAP}$: Camera-aware contrastive loss
  • $L_{KL}$: KL divergence loss between Teacher and Student
  • $L_{stri}$: Soft triplet loss on features

3. Role of Pseudo-Labeling and Refinement

• IQAGA: Does not use pseudo-labels in its training; adaptation is achieved exclusively by GAN-based augmentation and IQA reweighting.
• DAPRH: Applies clustering (DBSCAN) to infer hard pseudo-labels, which are then refined using cluster soft probabilities $G_i$ (based on distance to cluster centers), silhouette scores for filtering, and label smoothing: $\hat{y}_i = (1 - \alpha) y_i + \alpha G_i$. The pseudo-label refinement is iteratively applied each epoch, with a Teacher-Student model facilitating stable target learning.

4. Integration of Domain-Invariant and Holistic Feature Modules

• IQAGA: Does not incorporate domain-invariant mapping or holistic features.
• DAPRH: Uses a domain classifier $D_{net}$ for adversarial domain confusion during source training. During target training, features are enhanced via ViT-based holistic representations, incorporating both global and local spatial tokens. Further, CAP (camera-aware proxies) decomposes clusters by camera, and contrastive losses are employed at both cluster and camera levels.

5. Training Protocols and Hyperparameters

IQAGA

• GAN Augmentation: StarGAN trained with Adam ($lr=3.5 \times 10^{-5}$, $\lambda_{cls}=1$, $\lambda_{rec}=10$, $\lambda_{idt}=1$, $\lambda_{pid}=10$), batch size 16.
• Supervised Training: ResNet-50 pretrained on ImageNet, Adam ($lr=3.5 \times 10^{-4}$, reduced by 0.1 at epochs 40,70), batch size 128 (16 IDs × 8 images), 120 epochs, triplet margin $\alpha=0.3$, IQA reweighting $A_z=0.8$.

DAPRH

• GAN Augmentation: Same as IQAGA.
• Source & DIM: Batch size 128 (16 IDs × 8 images), $\lambda_{DIM}=0.1$.
• Target Training: DBSCAN $\epsilon=0.6$ (Market), minPts=8 (Market), 16 (MSMT), top-K local tokens $K=5$ (top 40%), $\alpha=0.4$, $\gamma=1.0$, $\beta_{KL}=0.4$, $\beta_{tri}=0.8$, EMA momentum $w=0.99$, 50–80 epochs, SGD $lr=1 \times 10^{-3}$.

6. Empirical Results and Comparative Analysis

Below are reported results on key cross-domain benchmarks, measuring mean Average Precision (mAP) and Rank-1 accuracy.

Method/Stage	Market→Duke (mAP / R1)	Duke→Market (mAP / R1)	Market→MSMT (mAP / R1)	Duke→MSMT (mAP / R1)
Baseline (ce+triplet)	25.8% / 43.7%	26.2% / 55.3%	–	–
+GAN only	31.5% / 54.2%	35.1% / 68.6%	–	–
+GAN+IQA (IQAGA final)	32.1% / 55.5%	36.3% / 70.2%	–	–
+GAN+DIM	35.4% / 58.2%	–	–	–
DAPRH final	72.0% / 83.7%	85.9% / 94.4%	35.8% / 64.8%	36.0% / 65.5%

GAN augmentation alone yields strong improvements over the baseline, with IQA-driven weighting (IQAGA) giving further 1% mAP gains. In DAPRH, the inclusion of holistic features, camera-aware proxies, CRL, and EMA achieves major additional improvements (e.g., 72.0% mAP Market→Duke). DAPRH narrows the gap to fully supervised methods (where SOTA achieves 98%+ locally) and surpasses earlier UDA results, which capped at ≤40 mAP on MSMT and ≤80 mAP on Market/Duke (Pham et al., 4 Jan 2026).

7. Qualitative Insights and Contributions

• GAN augmentation provides the largest early-stage gains in both methods, while DIM is computationally cheaper but less impactful alone.
• Additive combination of holistic representations (EIR/ViT backbone), camera-aware proxies, CRL, and Teacher-Student learning provides +2–4 mAP gains independently, with their combination yielding a further improvement of +1–2 mAP.
• Key hyperparameters such as $\alpha$ (label smoothing), $\gamma$ (CAP loss weight), and top-K local ratio exhibit optimal ranges: $\alpha \approx 0.4$–$0.6$, $\gamma \approx 1.0$, and top-K $\approx 0.4$.
• IQAGA demonstrates how simple GAN+IQA weighting yields a 32–36% mAP on Market→Duke, outperforming previous GAN-based UDA by 1–3 mAP. DAPRH, through integration of style-transfer, domain-alignment, advanced feature compositions, and robust pseudo-label refinement, achieves >70 mAP on Market→Duke and >85 mAP on Duke→Market (Pham et al., 4 Jan 2026).

The performance and methodology of IQAGA and DAPRH highlight that combining generative augmentation, domain-aligned representation learning, and sophisticated pseudo-labeling can bridge much of the cross-domain gap for ReID without requiring target-domain labels.