IQAGA and DAPRH: UDA for Person ReID

• IQAGA and DAPRH are advanced unsupervised domain adaptation frameworks that use GAN-based augmentation to bridge the gap between source and target person re-ID data.
• IQAGA employs a two-stage StarGAN-based style transfer with image quality weighting to mitigate artifacts and enhance training stability.
• DAPRH integrates domain-invariant mapping, cluster-based pseudo-label refinement, and holistic ViT features, achieving significant mAP improvements on standard benchmarks.

IQAGA (Image Quality–Driven GAN Augmentation) and DAPRH (GAN Augmentation + Pseudo-Label Refinement + Holistic Features) are two advanced unsupervised domain adaptation (UDA) frameworks developed to address cross-domain generalization in person re-identification (ReID) tasks where source and target data distributions diverge sharply due to appearance variation, camera-specific styles, and lack of target labels. Both approaches utilize generative adversarial networks (GANs) for domain-specific image augmentation but diverge in loss composition, feature engineering, domain-invariant mapping, and pseudo-label mechanisms. Evaluated on standard ReID benchmarks, these frameworks show major improvements over prior UDA methods by systematically integrating augmentation, feature supervision, and robust target data exploitation (Pham et al., 4 Jan 2026).

1. IQAGA: Image Quality–Driven GAN Augmentation

IQAGA centers on a two-stage workflow combining StarGAN-based style transfer and image-quality-weighted supervised learning. In Stage I, StarGAN models are trained to convert each source image $x_s$ into $C_t$ styles corresponding to target camera domains, optimizing multiple objectives: adversarial ($L_{adv}$), domain classification ($L_{cls}$), cycle-reconstruction ($L_{rec}$), identity mapping ($L_{idt}$), and identity-preserving color loss ($L_{pid}$). Each source image's translations are concatenated with the original source set to construct a synchronized training set $D_{sync}$.

In Stage II, ResNet-50 provides 2048-D features per image. Supervised learning optimizes cross-entropy ($L_{ce}$) and triplet ($L_{tri}$) losses, but with per-sample IQA weighting leveraging normalized feature vector statistics ($z_i$), directly modulating each image's gradient contribution: $L_i = (1 + A_z z_i)\cdot(L_{ce,i} + L_{tri,i})$. Low-quality GAN samples thus exert reduced influence, mitigating mode collapse and spurious artifacts.

Key design choices in IQAGA include avoidance of target pseudo-labeling and domain-invariant mapping: adaptation is driven purely by GAN-based augmentation and image-level loss engineering.

2. DAPRH: GAN Augmentation, Pseudo-Label Refinement, Holistic Features

DAPRH expands the GAN augmentation paradigm by combining (i) a domain-invariant mapping (DIM) adversarial feature alignment, (ii) cluster-based pseudo-labeling with refinement, (iii) holistic feature encoding via Vision Transformer (ViT), and (iv) camera-aware proxies.

Stage I uses StarGAN for style transfer as in IQAGA, but batch construction incorporates a reduced ratio (e.g., 4:1 real:GAN), curbing GAN noise. DIM employs a domain classifier $D_{net}$ to adversarially confound domain identity, training $f_e$ to push $D_{net}$ predictions toward 0.5 for both source and target features ($L_{DIM}$).

In Stage II, target images are encoded, clustered by DBSCAN, and assigned hard pseudo-labels $y_i$. Features are transformed by ViT/MLP, merging global (CLS token) and local MaxPool-selected top-K tokens for expressive queries $v_i$. Cluster centers $m_k$ support soft label refinement $G_i$ (via softmax on Euclidean distance), further filtered by silhouette coefficient $s_i$. Refined labels take the weighted form $\hat{y}_i = (1-\alpha)y_i + \alpha G_i$. A teacher-student framework employs EMA to update teacher weights, with student supervision via KL divergence ($L_{KL}$) and soft triplet ($L_{stri}$) losses.

DAPRH integrates camera-aware proxies by subdividing clusters via camera ID, yielding sub-centers $c_{k,b}$ and associated contrastive loss $L_{CAP}$ over proxies per sample. The unsupervised target-objective aggregates NCE, CAP, KL, and soft-triplet losses.

3. Mathematical Formulations

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:

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

• IQA-weighted Source Loss:

$$L_{source} = \sum_i \left[ (1 + A_z z_i) (L_{ce,i} + L_{tri,i}) \right]$$

DAPRH Losses (Stage I and II)

• DIM Loss:

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

• ClusterNCE:

$$L_{NCE} = -\log\left(\frac{\exp(q \cdot c^+/\tau)}{\sum_k \exp(q \cdot c_k/\tau)}\right)$$

• CAP Loss:

$$L_{CAP} = -\frac{1}{|P(i)|}\sum_{p \in P(i)} \log\left(\frac{\exp(q \cdot c_p/\tau_c)}{\sum_k \exp(q \cdot c_k/\tau_c)}\right)$$

• Pseudo-label refinement:

$$\hat{y}_i = (1-\alpha) y_i + \alpha G_i,\quad G_i[k] \propto \exp(-d_E(f_i, m_k)/t)$$

• Teacher EMA update:

$\theta_t \leftarrow w \theta_t + (1-w) \theta_s$

4. Training Protocols and Hyperparameterization

For both methods, StarGAN is trained on source versus target domain camera labels with Adam optimizer, learning rate $3.5\times10^{-5}$, batch size $16$, and loss weights $\lambda_{cls}=1$, $\lambda_{rec}=10$, $\lambda_{idt}=1$, $\lambda_{pid}=10$. ResNet-50 is initialized from ImageNet weights, supervised with Adam optimizer at $3.5\times10^{-4}$ (decayed at epochs $40,70$), batch size $128$, and $120$ epochs. Triplet margin $\alpha = 0.3$, IQA weight $A_z = 0.8$.

DAPRH source batch formation reserves $N:M$ real:GAN images (e.g. $4:1$), uses $\lambda_{DIM}=0.1$ for adversarial loss. Target clustering employs DBSCAN ($\epsilon=0.6$, MinPts=$8$ for Market, $16$ for MSMT), batch $128$. Top-K local tokens $K=5$ (approx 40% of total), $\alpha=0.4$, $\gamma=1.0$, $\beta_{KL}=0.4$, $\beta_{tri}=0.8$. Teacher EMA momentum $w=0.99$, $50$–$80$ epochs, SGD learning rate $1\times10^{-3}$.

5. Experimental Results and Ablation Findings

Quantitative results illustrate substantial gains in cross-domain scenarios:

Scenario	Baseline mAP / Rank-1	+GAN	+GAN+IQA / +DIM	DAPRH Final
Market→Duke	25.8 / 43.7	31.5 / 54.2	32.1 / 55.5	72.0 / 83.7
Duke→Market	26.2 / 55.3	35.1 / 68.6	36.3 / 70.2	85.9 / 94.4
Market→MSMT	—	—	—	35.8 / 64.8
Duke→MSMT	—	—	—	36.0 / 65.5

Ablation analyses show:

• GAN augmentation alone yields $>10$ mAP improvement; IQA weighting adds another $\sim1$ mAP in IQAGA.
• In DAPRH, DIM is more computationally efficient than GAN for early-stage feature alignment, but integrating both is optimal.
• Holistic (ViT) features and CAP each contribute $+2$–$4$ mAP; jointly, a further $+1$–$2$ mAP.
• CRL and teacher-student boost $+1$–$2$ mAP, crucial for scaling to large datasets.
• Key hyperparameters exhibit clear optima: $\alpha \approx 0.4$–$0.6$, $\gamma \approx 1.0$, top-K $\approx 0.4$.

This suggests that high-fidelity augmentation, loss weighting, domain confusion, and advanced pseudo-labeling together address critical bottlenecks in fully unsupervised ReID adaptation.

6. Contributions and Comparative Significance

IQAGA demonstrates that simple GAN-based augmentation, when augmented with IQA-driven sample weighting, surpasses prior GAN-based UDA by $1$–$3$ mAP. DAPRH incorporates multi-component alignment—style transfer, adversarial mapping, refined soft pseudo-labels, holistic feature representation, camera-aware local proxies—and achieves more than $70$ mAP on Market→Duke and $85$ mAP on Duke→Market, bridging much of the practical gap to fully supervised approaches. On large-scale MSMT, DAPRH exceeds $40$ mAP, surpassing previous unsupervised adaptation results.

A plausible implication is that multi-stage integration of style transfer, discriminative feature enhancement, cluster-based label refinement, and domain-invariant mapping forms an effective paradigm for cross-domain ReID without target labels. Further, the critical role of image quality assessment and proxy learning highlights the importance of robust sample and feature selection in deep UDA pipelines, a point of emerging significance for unsupervised visual recognition research (Pham et al., 4 Jan 2026).