Semi-Supervised Instance Segmentation

• The paper introduces a framework that uses teacher–student self-training to generate reliable pseudo-labels from limited annotated data.
• It employs adaptive weighting, score thresholding, and consistency regularization to ensure high-quality segmentation across diverse modalities.
• Empirical findings demonstrate state-of-the-art improvements in mask accuracy and efficiency, reducing the burden of dense annotations.

A semi-supervised instance segmentation framework addresses the challenge of learning to delineate and classify object instances with incomplete supervision—that is, from a combination of a small labeled set and a large pool of unlabeled (or weakly labeled) data. Motivations for such frameworks arise from the prohibitive cost of dense pixel-wise annotations required by traditional supervised pipelines, especially in applications such as biomedical imaging, 3D scene understanding, and industrial or medical analysis. Modern semi-supervised instance segmentation leverages recent advances in teacher–student self-training, consistency regularization, pseudo-labeling, contrastive learning, and data-centric augmentation, and can be broadly instantiated for both 2D and 3D modalities. The following sections synthesize the main algorithmic patterns, supervision protocols, architectural components, and key empirical findings across recent literature.

1. Teacher–Student Self-Training Paradigms

A canonical methodological axis for semi-supervised instance segmentation is the teacher–student (or Mean Teacher) paradigm (Zhou et al., 2020, Filipiak et al., 2022, Hu et al., 2023, Yoon et al., 7 Apr 2025, Taghavi et al., 28 May 2025). Here, two models with identical (or similar) architectures are initialized and updated in parallel: the teacher is typically synchronized to the student via exponential moving average (EMA) of its weights. Labeled images are used for standard supervised losses; on unlabeled data, the teacher generates pseudo-labels which serve as training targets for the student after appropriate filtering or reweighting. Perturbations or data augmentations (weak/strong) are crucial to prevent shortcut learning and enforce robustness via consistency regularization—requiring the student to output consistent predictions even under diverse appearance transformations (Zhou et al., 2020, Filipiak et al., 2022).

In 3D settings, such as SPIB (Liao et al., 2021) or consistency-regularized approaches for panoptic segmentation (Coenen et al., 2022), the teacher–student loop is adapted using spatial perturbations, point-level transformations, and unlabeled geometric data streams. Pseudo-labels (boxes, masks, or orientation/distance transforms) are generated by the teacher and drive both supervised and unsupervised objectives for the student.

2. Pseudo-Label Generation and Quality Control

Pseudo-labeling—the process of generating artificial labels for unlabeled data—underpins most modern semi-supervised frameworks. The challenge lies in filtering noisy or incorrect pseudo-labels, which can lead to confirmation bias and error accumulation during self-training. Various filtering strategies are employed:

• Score thresholding and mask scoring: Mask R-CNN and CenterMask-based pipelines (Filipiak et al., 2022, Wang et al., 2022) compute per-instance class and mask-quality scores; only labels exceeding confident thresholds (e.g., $\tau_\mathrm{cls}$ for classification, $\tau_\mathrm{IoU}$ for mask quality) are retained. Masks below thresholds are ignored or used as hard negatives for classification (Filipiak et al., 2022, Hu et al., 2023).
• Adaptive or dynamic weighting: In PAIS (Hu et al., 2023), the losses associated with pseudo-instances are re-weighted as a function of their predicted confidence via the Dynamic Aligning Loss (DALoss), rather than discarded. This approach maintains signal from "borderline" pseudo-labels, preventing over-pruning and leveraging instances with high segmentation fidelity but low class confidence, or vice versa.
• Self-ensemble pseudo-labels: Mean Teacher/EMA paradigms (Zhou et al., 2020, Filipiak et al., 2022) or ensembling over multiple augmentations mitigate noise by averaging predictions from several teacher passes (with different perturbations), followed by entropy or sharpness calibration (Zhou et al., 2020).
• Instance- or class-specific filtering: In domain-specific applications, additional pseudo-label filtering is applied at the instance-class level (e.g., per-class mask thresholding for tiny structures in StomataSeg (Huang et al., 31 Jan 2026)) or via domain priors such as occupancy ratios in 3D point clouds (Liao et al., 2021).

3. Semi-Supervised Losses and Consistency Regularization

Frameworks employ joint loss functions combining supervised, pseudo-supervised, and consistency terms. The structure typically follows:

$$\mathcal{L} = \mathcal{L}_{\mathrm{sup}} + \lambda_u \mathcal{L}_{\mathrm{unsup}} + \lambda_\mathrm{cons}\mathcal{L}_{\mathrm{cons}},$$

where:

• Supervised loss ($\mathcal{L}_{\mathrm{sup}}$): Standard objectives (e.g., cross-entropy, Dice, IoU) computed on labeled data and masks.
• Pseudo-label loss ($\mathcal{L}_{\mathrm{unsup}}$): Pseudo-labeled images are treated as labeled but may receive lower or dynamic weights (Bellver et al., 2019, Huang et al., 31 Jan 2026).
• Consistency loss ($\mathcal{L}_{\mathrm{cons}}$): Regularization enforcing agreement under perturbation (spatial, photometric) or structural alignment between teacher and student. For instance, CAST (Taghavi et al., 28 May 2025) employs an instance-aware contrastive loss to pull together pixel embeddings within instances and push apart others, while ConsInstancy (Coenen et al., 2022) aligns semantic and instance-oriented representations.

Feature-level distillation, as in Mask-guided Mean Teacher (MMT-PSM) (Zhou et al., 2020), further restricts the feature consistency term to foreground regions using mask-derived spatial priors.

4. Architectural Adaptations and Backbones

Architectural choices reflect both application domains and the supervision regime:

• Detection-based instance segmentation: Most semi-supervised pipelines build on Mask R-CNN or its variants (with enhancements such as feature pyramid networks, mask scoring heads, or anchor-free detectors like CenterMask2 (Filipiak et al., 2022)), ensuring high recall of initial proposals.
• Specialized decoders and heads: To handle noisy labels, architectures may introduce low-resolution mask branches for boundary noise suppression (Noisy Boundaries (Wang et al., 2022)), confidence prediction heads (e.g., box- and mask-IoU heads in (Wang et al., 2020)), or use orientation/distance fields for panoptic consistency (Coenen et al., 2022).
• Domain-specific adaptations: 3D instance segmentation frameworks for point clouds or medical images utilize specialized feature extractors (e.g., VoteNet (Liao et al., 2021)), coarse-to-fine cascades (as in dental CBCT (Wang et al., 28 Nov 2025)), or fusion with auxiliary modalities, as in depth-guided instance segmentation (Chen et al., 2024).

■■ Table: Sample Semi-supervised Instance Segmentation Architectures

Paper/Framework	Backbone/Network	Special Modules/Heads
MMT-PSM (Zhou et al., 2020)	Mask R-CNN + IR-Net	Perturb-sensitive mining, mask-guided distillation
PAIS (Hu et al., 2023)	Mask R-CNN, K-Net	Dynamic Aligning Loss, mask-IoU head
CAST (Taghavi et al., 28 May 2025)	DINOv2-S/Mask2Former	Instance-aware pixel contrastive loss, multi-stage distillation
StomataSeg (Huang et al., 31 Jan 2026)	Mask R-CNN, ConvNeXt-V2	Patch cropping, per-class pseudo filtering
ConsInstancy (Coenen et al., 2022)	FCN/RS-Net	Orientation/distance maps, panoptic postproc
S⁴M (Yoon et al., 7 Apr 2025)	Mask2Former, SAM	SAM-based mask refinement, ARP augmentation

5. Data and Annotation Protocols

Annotation budgets and protocols are pivotal design axes in semi-supervised instance segmentation:

• Budget-aware supervision: Explicit modeling of annotation time enables empirical budget–performance curves (e.g., 239.7s for full-mask, 38.1s for boxes, 23.3s for points per image in VOC (Bellver et al., 2019, Kim et al., 2023)), informing optimal allocation of strong (mask), weak (point/box/image-level), and unlabeled supervision.
• Hybrid weak/semi-supervision: Protocols such as weakly semi-supervised instance segmentation with point labels (WSSIS) (Kim et al., 2023) use a small strong label set with large point-labeled pools, enabling highly efficient pseudo-labeling that avoids heavy reliance on confidence thresholds.
• Dataset handling: Splitting high-resolution images into overlapping patches, as in StomataSeg (Huang et al., 31 Jan 2026), addresses the small-object regime by increasing effective annotation and pseudo-label coverage for objects under 40 μm.

6. Empirical Results and Comparative Insights

Extensive benchmarking confirms consistent superiority of semi-supervised methods over strong-only or weakly supervised baselines—even when labeled data is as low as 1–5%. Recent advances yield state-of-the-art mask AP (mAP) and region/instance affinity metrics on datasets such as COCO, Cityscapes, BDD100K, and custom biomedical datasets.

■■ Table: Comparative Results (Mask AP / mAP or Comparable Metrics)

Method	COCO (1%)	COCO (5%)	Cityscapes (5%)	Notes
Supervised baseline	3.5	17.3	12.10	(respective class/mask AP)
NoisyBoundaries (Wang et al., 2022)	7.7	24.9	17.1
PAIS + Mask R-CNN (Hu et al., 2023)	21.2	29.3	18.0	DALoss weighting
S⁴M (Ours) (Yoon et al., 7 Apr 2025)	24.2	32.1	30.1	SAM-based mask refinement
Depth-Guided (Chen et al., 2024)	22.3	31.5	23.2	DPT-based depth fusion
StomataSeg (Huang et al., 31 Jan 2026)	–	–	–	+21 AP gain for fine objects

Performance gains are attributed to careful pseudo-label curation, architectural regularization, advanced augmentation, and hybrid domain fusion (e.g., depth (Chen et al., 2024), SAM guidance (Yoon et al., 7 Apr 2025), or layout/shape priors (Chen et al., 2023)). Notably, S⁴M achieves +6.9 mAP improvement at 5% labeled data in Cityscapes over NoisyBoundaries. In the medical imaging domain, semi-supervised and pseudo-labeling strategies boosted instance affinity and Dice scores by more than 44–61 points over baselines (Wang et al., 28 Nov 2025).

7. Current Limitations and Future Directions

While semi-supervised frameworks have drastically reduced the annotation burden, several open challenges persist:

• Pseudo-label selection vs. weighting: Methods such as PAIS suggest that soft weighting (DALoss) of potentially noisy pseudo-labels outperforms hard filtering at low thresholds, but optimal tuning per dataset/domain remains open.
• Domain adaptation and generalization: The utility of foundational models such as SAM depends on careful adaptation (e.g., box-prompt engineering, normalization strategies) for medical or low-resource domains (Yoon et al., 7 Apr 2025, Wang et al., 28 Nov 2025).
• 3D/temporal and small-object regimes: Fine structures (e.g., plant stomata (Huang et al., 31 Jan 2026), aggregate particles (Coenen et al., 2022)), and temporally coherent video frameworks (Le et al., 2021) demand tailored preprocessing, patchification, synthetic augmentation, and spatial/temporal consistency modules.
• Extension to multi-modal and active learning: Fusion with auxiliary signals (depth (Chen et al., 2024), skeletons (Le et al., 2021)), or incorporation of active querying strategies (TSP (Wang et al., 2020)) are active research avenues.
• Complexity and computational load: Multi-stage cascades, foundation model integration, and heavy post-processing (e.g., in dentistry and 3D microscopy) require substantial computational and engineering investment (Wang et al., 28 Nov 2025, Yoon et al., 7 Apr 2025).

Research directions include automated confidence calibration for pseudo-label selection, more expressive shape or geometric priors, hierarchical consistency (semantic, instance, panoptic), and modular transfer of foundational vision models to constrained or emerging domains. The field is converging on standardized self-training protocols with rigorous ablation, interpretability analyses, and open-source reproducibility.

---

References:

(Zhou et al., 2020, Liao et al., 2021, Kim et al., 2023, Chen et al., 2024, Filipiak et al., 2022, Huang et al., 31 Jan 2026, Yoon et al., 7 Apr 2025, Hu et al., 2023, Coenen et al., 2022, Wang et al., 2020, Taghavi et al., 28 May 2025, Wang et al., 2022, Bellver et al., 2019, Wang et al., 28 Nov 2025, Le et al., 2021, Chen et al., 2023)