Class-Aware Sampling

• Class-aware sampling is a strategy that uses label information to adjust sample frequency, weighting, or selection to better represent diverse classes.
• It improves model performance in imbalanced tasks such as long-tailed detection, meta-learning, and federated learning by prioritizing difficult and rare examples.
• Empirical studies show that class-aware methods enhance metrics like PSNR in denoising and mAP in detection while mitigating overfitting risks.

Class-aware sampling denotes any data selection strategy that incorporates label class information to guide the choice, frequency, or weighting of samples presented to a learning algorithm. In contrast to class-agnostic (uniform or random) sampling, class-aware schemes explicitly adapt to the class distribution, complexity, or semantic prior, often with the objective of improving model generalization, representation of rare classes, or preservation of domain-specific structure. Recent work employs class-aware sampling in diverse settings including image denoising (Remez et al., 2017), meta-learning (Liu et al., 2020), federated learning (Zhang et al., 2022), long-tailed detection (Yaman et al., 2023, Ahmed et al., 27 Mar 2025), dataset quantization (Zhao et al., 9 Jul 2024), and segmentation (Huang et al., 2022).

1. Core Principles and Taxonomy

Class-aware sampling algorithms can be grouped according to their operational mechanism and intended objective:

• Uniform Class Rebalancing: Equalizes per-class sample frequency via oversampling minority classes or undersampling majority classes.
• Difficulty-Driven or Hardness-Aware Sampling: Prioritizes samples belonging to classes or instances identified as more challenging for the current classifier, regularly updating based on error feedback (Chen et al., 2018, Mo et al., 2020, Liu et al., 2020).
• Instance-Aware and Feature-Space Sampling: Considers both image and instance counts, or leverages embedding space proximity and informativeness for sample selection (Yaman et al., 2023, Zhao et al., 9 Jul 2024).
• Adaptive Sampling and Quantization: Dynamically allocates more samples to sensitive classes which benefit from incremental data, reducing allocation to stable classes with saturated accuracy (Zhao et al., 9 Jul 2024).
• Task- and Trajectory-Class Awareness: In meta-learning and RL, tasks or episodes are constructed to maximize difficulty or diversity based on class-pair potentials or trajectory clustering (Liu et al., 2020, Na et al., 3 Mar 2025).
• Topology- and Shape-Aware Approaches: Uses domain-specific measures (fractal dimension, skeletonization) to allocate sampling and loss weighting according to structural complexity (Zhang et al., 14 Jun 2025).

These principles are operationalized in both static data pipelines (long-tailed detection, segmentation) and active or iterative settings (meta-learning, federated systems, dataset distillation).

2. Mathematical Formulations

Representative mathematical models from recent works:

• Repeat Factor Sampling (RFS) and IRFS: Class-level repeat factor:

$$r_c = \max\left(1, \sqrt{\frac{t}{\sqrt{f_{i, c} f_{b, c}}}} \right)$$

where $f_{i, c}$ and $f_{b, c}$ are the per-class image and instance frequencies (Yaman et al., 2023), $t$ is a threshold.

• Exponentially Weighted IRFS (E-IRFS):

$$r_c = \exp\left(\alpha \sqrt{\frac{t}{\sqrt{f_{i, c} f_{b, c}}}} \right)$$

amplifies rare-class upsampling (Ahmed et al., 27 Mar 2025).

• SampleAhead distribution update:

$$P^{(t)}(U) \propto \alpha P^{(0)}(U) + (1 - \alpha) P^{(0)}(U) \exp(\beta d^{(t-1)}(U))$$

where $d^{(t-1)}(U)$ is the estimated sample difficulty (Chen et al., 2018).

• Fed-CBS QCID (Quadratic Class-Imbalance Degree):

$$\mathrm{QCID}(\mathcal{M}) = \sum_{b=1}^B \left( \frac{ \sum_{n \in \mathcal{M}} q_n \alpha_{n, b} }{ \sum_{n \in \mathcal{M}} q_n } - \frac{1}{B} \right)^2$$

measures imbalance in federated selection (Zhang et al., 2022).

• Shape-Aware Patchsize Allocation (FDPS):

$$\mathrm{FDPS}_i = \text{Assign}(\mathrm{PS}_{\text{min}}, \mathrm{PS}_{\text{mid}}, \mathrm{PS}_{\text{max}} | \mathrm{FD}_{i}= \mathrm{FD}_{\text{max}}, \mathrm{FD}_{\text{mid}}, \mathrm{FD}_{\text{min}})$$

matches patchsize to axis complexity (Zhang et al., 14 Jun 2025).

• Meta-Learning Class-Pair Potentials:

$$p_{CP}^{(t+1)}(S_K) \propto \prod_{i, j \in S_K} C^{(t)}(i, j)$$

$C^{(t)}(i, j)$ is iteratively updated by episode misclassification statistics (Liu et al., 2020).

3. Case Studies and Empirical Evidence

• Image Denoising: Class-aware fine-tuning improves PSNR up to $0.4$ dB over class-agnostic baselines, with denoisers learning class-specific priors (e.g. facial detail in portraits) (Remez et al., 2017).
• Long-Tailed Detection: IRFS and E-IRFS markedly outperform RFS in rare-class AP, with E-IRFS producing up to $350\%$ mAP$_{50}$ improvement for the rarest classes in UAV detection (Yaman et al., 2023, Ahmed et al., 27 Mar 2025).
• Active Learning for Detection: Box-level class-balanced sampling, weighting uncertainty by class frequency, yields up to $2.8\%$ mAP improvement at low annotation budgets and enhances minority-class detection (Liao et al., 25 Aug 2025).
• Few-shot Segmentation: Hybrid class-aware and class-agnostic alignment improved mIoU and foreground-background IoU in challenging settings (notably, 1-shot scenarios) (Huang et al., 2022).
• Federated Learning: Fed-CBS class-balanced client selection achieves convergence $1.3\times$ to $2.8\times$ faster than state-of-the-art, outperforming random client selection and attaining high accuracy even under severe heterogeneity (Zhang et al., 2022).
• Adaptive Latent Space Sampling: Hardness-aware strategies concentrate training on ambiguous areas near class boundaries, achieving higher accuracy than uniform sampling in MNIST and CIFAR-10 (Mo et al., 2020).
• Dataset Quantization: DQAS adaptively reduces data for stable classes where accuracy is saturated, increasing it for sensitive classes, resulting in superior performance at aggressive compression ratios (Zhao et al., 9 Jul 2024).

4. Implementation Strategies and Technical Considerations

• Pipeline Integration: Class-aware sampling is typically implemented as a data loader or sampling wrapper (repeat factor, per-class binning, sampling from class-specific buckets). Extensions include feature-space binning and skeleton-aware loss integration.
• Adaptivity: Several methods (SampleAhead, DQAS) include iterative feedback between model performance and sampling distribution. Bucket-based approximations or heuristics are used for computational efficiency.
• Representation and Domain Priors: Some approaches utilize external knowledge—e.g., text-prompted diffusion models for anatomical structure synthesis (Varaganti et al., 29 Apr 2025), meta-learning class-pair potentials for episode sampling (Liu et al., 2020), or trajectory clustering for multi-agent RL (Na et al., 3 Mar 2025).
• Risk Management: Over-specialization or overfitting risks are mitigated via augmentation (Breadcrumbs/EMANATE (Liu et al., 2021)), adversarial feature trails, or balancing hard/easy sample selection with importance weights (Grosz et al., 20 Sep 2024).
• Hyperparameter Tuning: Methods often control aggressiveness through scaling parameters (e.g., $\alpha$ in E-IRFS) and thresholds, requiring domain-specific calibration.

5. Comparative Analysis and Limitations

Sampling Method	Class Awareness	Strengths	Common Limitations
RFS	Image-level	Simple, plug-and-play	Ignores instance counts
IRFS	Image+Instance	Substantially better rare-class recall	Still linear scaling, limited for ultra-rare classes
E-IRFS	Exp. scaling	Superior in extreme imbalance	Must tune exponential scaling
Breadcrumbs/EMANATE	Feature-trail	Mitigates few-shot overfitting	Assumes access to prior epochs
SampleAhead	Difficulty-driven	Efficient in infinite/synthetic space	Requires accurate difficulty estimation
Meta-Learning Task	Class-pair/taks	Improves generalization, explores hard class-pairs	Computational for large episode sets

Potential limitations include:

• Necessity of precise class label partitions or reliable classifier priors (ambiguity in class assignment reduces benefit).
• Added pipeline complexity (sample selection, classifier integration, augmentation scheduling).
• Risk of overfitting in specialized or oversampled classes, especially when using naive oversampling rather than augmentation or synthetic generation.
• Implementation and computational cost for dynamic or iterative feedback-based methods.

6. Application Domains and Broader Implications

Class-aware sampling has demonstrated impact in several domains:

• Low-level vision: Denoising, super-resolution, inpainting, deblurring with class- or domain-specific priors (Remez et al., 2017).
• Imbalanced recognition: Real-world detection for autonomous driving, surveillance, medical imaging (CT, ultrasound, anatomical landmark detection) (Yaman et al., 2023, Varaganti et al., 29 Apr 2025).
• Meta-learning: Few-shot learning, episodic training curricula, class-pair curriculum design (Liu et al., 2020).
• Federated learning: Fairness and convergence under non-IID client data (Zhang et al., 2022).
• Quantized/Distilled datasets: Compressed training via class-sensitive selection reduces annotation and computation costs (Zhao et al., 9 Jul 2024).
• Multi-agent RL: Trajectory-class-aware policies and joint behavior adjustment in multi-task settings (Na et al., 3 Mar 2025).
• Topology-sensitive segmentation: Medical or remote sensing tasks where structure preservation is critical (airways, vessels, roads) (Zhang et al., 14 Jun 2025).

Class-aware sampling represents a scalable, general paradigm for enhancing learning in domains where class imbalance, complexity, or semantic specificity affects downstream performance.

7. Future Directions and Controversies

Ongoing research explores:

• Adaptive and dynamic parameterization: Online adjustment of class-aware sampling factors (e.g., dynamic adaptation of exponential scaling).
• Integration with self-supervision: Hybrid sampling strategies leveraging both labeled and unlabeled domains.
• Extension to new modalities: Further generalization into audio, text, and graph data, with class-aware sampling tailored to domain structure.
• Combination with augmentation and synthetic data: Combining sampling with generative models for minority class enrichment (Varaganti et al., 29 Apr 2025).
• Risk minimization: Techniques to avoid overfitting, including adversarial augmentation, soft pseudo labeling, and informed instance weighting (Liu et al., 2021, Grosz et al., 20 Sep 2024).

A plausible implication is that class-aware sampling, when judiciously combined with adaptive augmentation and error-driven update mechanisms, will continue to deliver efficiency and fairness improvements in increasingly complex, multi-modal learning systems. However, appropriate tuning and domain adaptation, and strategies to avoid over-specialization, remain central to its success.