Three-Stage Pseudo-Label Curation

• Three-stage pseudo-label curation is a structured pipeline that generates, filters, and refines labels to mitigate noise in weakly supervised learning scenarios.
• The pipeline employs advanced techniques such as confidence thresholding, distance filtering, and iterative refinement to enhance label quality across diverse tasks.
• Empirical results demonstrate significant performance gains and reduced annotation overhead, achieving near supervised performance in areas like biomedical segmentation and domain adaptation.

A three-stage pseudo-label curation pipeline is a structured methodology for generating, filtering, and refining pseudo-labels in weakly supervised, noisy, or domain-adaptive machine learning scenarios. This paradigm has emerged as a critical tool for reducing the need for full manual annotation, especially in settings such as biomedical image segmentation, domain adaptation, few-shot and weakly supervised learning, and temporal action localization. Pipelines in this category orchestrate three distinct operations—initial pseudo-label generation, quality-driven filtering or reweighting, and targeted refinement or human-in-the-loop correction—resulting in substantial performance gains and annotation savings compared to naive pseudo-labeling approaches (Zhao et al., 6 Nov 2025, Chhabra et al., 9 Feb 2024, Huang et al., 2023, Feng et al., 12 Jul 2024).

1. Overview and Common Structure

Three-stage pseudo-label curation pipelines generally follow this operational structure:

1. Automated Pseudo-Label Generation: Application of a pre-trained model, clustering, manifold propagation, or other automated method to assign tentative labels to unlabeled data. These pseudo-labels are usually high-coverage but noisy.
2. Curation via Selection, Filtering, or Reweighting: Application of algorithmic filters—often based on confidence, feature-space distance, class balance, self-consistency, or error likelihood—to select, weight, or modify pseudo-labels for maximizing label quality and preventing propagation of label noise.
3. Refinement through Human or Model-driven Correction: Further improvement of labels by expert correction (often in a strategically selected subset), iterative model-based re-labeling, or semi-automatic retraining, aimed at maximal data efficiency and robust generalization.

Iterative execution across these stages, sometimes with explicit convergence criteria or budget constraints, is a recurring pattern (Zhao et al., 6 Nov 2025, Lazarou et al., 2020).

2. Key Methodologies and Implementation Variants

2.1 Pseudo-Label Generation

Initial pseudo-labels are generated by:

• Running a foundation or pre-trained model on each data instance (e.g., cellSAM or SAM for images (Zhao et al., 6 Nov 2025, Huang et al., 2023), segmenters, or LLMs (Asano et al., 18 Feb 2025)).
• Label propagation via k-nearest-neighbor manifold graphs for few-shot tasks (Lazarou et al., 2020).
• Gaussian Mixture Model–based classifiers for domain adaptation, using model-internal posteriors (Chhabra et al., 9 Feb 2024).
• Cross-video contrastive mining in temporal action localization (Feng et al., 12 Jul 2024).

This stage prioritizes recall and global coverage, accepting that most labels require downstream curation.

2.2 Quality-driven Filtering and Curation

Mechanisms implemented in the second stage include:

• Confidence Thresholding: Only retaining instances exceeding a dynamically scheduled confidence (e.g., $p_{\tau}(t)$) (Chhabra et al., 9 Feb 2024, Rufin et al., 2023).
• Distance/Conformity Filtering: Retaining labels whose feature embeddings are close to class prototypes or cluster centers (e.g., $z_x \leq z_{\mathrm{th}}$) (Chhabra et al., 9 Feb 2024, Bui-Tran et al., 29 Oct 2025).
• Consistency Checking: Enforcing temporal or iterative agreement of labels across training epochs (Chhabra et al., 9 Feb 2024).
• Instance-level Metrics: Using proposal density (IoU-overlap), size, semantic confidence, and adaptive site- or instance-wise thresholds (Rufin et al., 2023, Feng et al., 12 Jul 2024).
• Class Balancing and Distribution Regularization: Sinkhorn-Knopp balancing of pseudo-labels to match target priors (Lazarou et al., 2020).
• Uncertainty-based Masking: Quantile-based entropy filters and prototype-based denoising (Bui-Tran et al., 29 Oct 2025, Huang et al., 2023).

Ablation studies consistently show that quality-driven selection in this stage delivers >20-point absolute performance improvements over naive pseudo-labeling in various benchmarks (Chhabra et al., 9 Feb 2024, Zhao et al., 6 Nov 2025).

2.3 Iterative Refinement, Correction, and Fine-Tuning

Curated labels undergo refinement via:

• Manual annotation of a feature-diverse core-set selected by farthest-first feature traversal (Zhao et al., 6 Nov 2025).
• Uncertainty-based self-correction using Monte Carlo dropout, correcting only pixels with model-predicted labels that differ from the initial pseudo-label and exhibit high confidence (Huang et al., 2023).
• Training/fine-tuning the model on the curated (and possibly partially relabeled) set, using losses that combine standard supervision, deviation-penalizing terms (e.g., Dice + cross-entropy), and explicit weighting to avoid overfitting to the noisy pseudo-labels.
• Model-based iterative relabeling, as in robust-unlabeled learning cycles (UU learning), with each classifier output used to re-define pseudo-positive/negative pools for the next round until class priors converge (Asano et al., 18 Feb 2025).
• Maintaining an Exponential Moving Average (EMA) teacher for distillation and smoothing of proposals (Feng et al., 12 Jul 2024, Bui-Tran et al., 29 Oct 2025).
• Mixup or data augmentation regularization to further stabilize training and suppress confirmation bias (Saravanan et al., 7 Feb 2024, Jo et al., 2022).

3. Application Areas and Domain Specific Extensions

3.1 Biomedical Image Segmentation

In microscopy and biomedical scenarios (e.g., MitoEM), the pipeline is instantiated by foundation model pseudo-labeling (cellSAM), automated pre-training of a fully-convolutional network (nnU-Net), and minimal-cost manual correction of feature-core-set patches via interactive tools (microSAM), yielding >90% of fully supervised performance at <13% of annotation cost (Zhao et al., 6 Nov 2025).

3.2 Domain Adaptation and Robust Classification

For unsupervised domain adaptation, pseudo-labels are scheduled and filtered through confidence, conformity, and consistency metrics, culminating in progressive inclusion for classifier retraining (Chhabra et al., 9 Feb 2024). In partial-label learning, weighted nearest-neighbor voting, label smoothing, and iterative partial-label expansion are key algorithmic steps (Saravanan et al., 7 Feb 2024).

3.3 Temporal Action Localization

FuSTAL employs cross-video contrastive proposal mining, prior-based density filtering, and EMA-distilled label refinement, obtaining mAP improvements >10% (from 40.9% to 50.8% on THUMOS’14) across the three stages (Feng et al., 12 Jul 2024).

3.4 Few-shot Learning

Manifold-based label propagation and Sinkhorn-based balancing of pseudo-labels are looped with loss-based selection of cleanest labels per class, with each iteration incrementally moving high-confidence queries into the support set (Lazarou et al., 2020).

3.5 Segmentation with Limited Annotation

Label correction frameworks for SAM-based segmentation rely upon pixel-level and image-level quality weighting, followed by uncertainty-guided relabeling and structured retraining, delivering Dice improvements of 2–4 percentage points over baselines and approaching supervised upper bounds (Huang et al., 2023).

4. Quantitative Impact and Empirical Findings

Across varied tasks and domains, the three-stage pseudo-label curation pipeline delivers pronounced gains:

Task/Domain	Key Metric	Baseline	Final (3-stage)	Label Cost Saving
Instance segmentation	F1, Panoptic, Precision	0.4025 (F1)	0.6003 (12.5% manual)	>90% of full perf. at <13% labels (Zhao et al., 6 Nov 2025)
Domain adaptation	SVHN→MNIST Accuracy	61.5%	98.9%	Unlabeled—all pseudo-label based (Chhabra et al., 9 Feb 2024)
Field delineation	mIoU, mRMSE (field size)	0.634 (mIoU)	0.674 (pseudo-only)	77% of human-label gain (Rufin et al., 2023)
Medical segmentation	Dice (JSRT, etc.)	88.5% (SAM)	91.88% (MLC pipeline)	No expert annotation required (Huang et al., 2023)
WSTAL (video)	mAP on THUMOS’14	40.9%	50.8%	Full weak-supervision (video-level labels) (Feng et al., 12 Jul 2024)

Ablation studies demonstrate that removing any curation/filtering/refinement stage consistently leads to nontrivial performance drops (up to 4–10 percentage points depending on the dataset), confirming the additive importance of each operation (Zhao et al., 6 Nov 2025, Feng et al., 12 Jul 2024, Huang et al., 2023).

5. Methodological Best Practices and Implementation Recommendations

• Feature Diversity: Core-set selection based on latent feature coverage is superior to random or image-level sampling.
• Adaptive Thresholds: Site-/sample-specific confidence or percentile cutoffs (rather than global thresholds) yield superior pseudo-label sets (Rufin et al., 2023).
• Iterative Filtering: Stage-wise rejection or reweighting based on label stability, conformity, and entropy is crucial for minimizing label noise and confirmation bias.
• Self-supervised Backbones: Self-supervised representations for feature extraction (e.g., masked autoencoders) enhance core-set coverage and pseudo-label quality (Zhao et al., 6 Nov 2025).
• Limited Human Oversight: Strategically allocate annotation effort to feature-representative samples, using interactive tools to minimize time.
• Cross-validation: Always validate pseudo-label performance against a small hold-out of human-annotated data to detect systematic biases.
• Iterative Convergence: In cycles where pseudo-labels are iteratively redefined, monitor label distributions or validation metrics to preempt overfitting or error propagation (Asano et al., 18 Feb 2025, Lazarou et al., 2020).

6. Limitations and Prospects

Three-stage pseudo-label curation pipelines, while transformative in reducing annotation burden, are contingent on the initial pseudo-label generator's generalization (i.e., foundation model or pre-trained backbone), the discriminative power of feature embeddings for selection, and the quality of uncertainty or entropy estimates used in curation. In extreme domain adaptation, or when feature distributions are highly multimodal or develop class collapse under weak supervision, additional refinement or more sophisticated filtering (e.g., entropy regularization, advanced clustering) may be required. Nevertheless, empirical evidence from biomedical (Zhao et al., 6 Nov 2025, Huang et al., 2023), agricultural (Rufin et al., 2023), and generic benchmarks (Chhabra et al., 9 Feb 2024, Feng et al., 12 Jul 2024) confirms the generalizability of this methodology across domains.

7. Relationship to Related Paradigms and Future Directions

These curation pipelines generalize and subsume classical self-training, active learning with core-sets, manifold label propagation, and semi-supervised learning architectures. Recent instantiations combine self- and cross-supervised learning, automated core-set extraction, uncertainty-based self-correction, and foundation model inference, moving toward fully automated, annotation-efficient pipelines. Future directions may include deeper integration with large-scale generative foundation models, more sophisticated uncertainty quantification, continual learning settings, and domain-agnostic selection mechanisms for pseudo-label refinement. Advances in interactive segmentation and self-supervised feature learning will continue to improve the efficiency and reliability of each stage in the curation pipeline (Zhao et al., 6 Nov 2025, Huang et al., 2023, Rufin et al., 2023).