Semi-supervised Instance Segmentation
Also known as: Semi-supervised Mask R-CNN, pseudo-label instance segmentation, label-efficient instance segmentation, SSIS
Semi-supervised instance segmentation trains a model to detect and delineate every object instance in an image using a small labeled set and a large unlabeled image corpus. By generating pseudo-labels from confident predictions on unlabeled images and enforcing consistency under augmentation, the approach achieves competitive mask accuracy at a fraction of the full annotation cost.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use semi-supervised instance segmentation when you need pixel-level object masks at scale but annotation budget allows only a small fraction of images to be labeled manually. Ideal for medical imaging, satellite imagery, and industrial inspection where expert labeling is costly. Works best when the unlabeled images share the same domain as the labeled set. Do not use when the labeled and unlabeled sets differ substantially in domain or object distribution, as pseudo-labels will be unreliable. Also avoid when the labeled set is below roughly 100 images per class — the base model will be too weak to generate useful pseudo-labels and fully supervised learning on a smaller, carefully curated labeled set may be preferable.
Strengths & limitations
- Dramatically reduces the number of pixel-level annotations needed for competitive mask accuracy.
- Leverages abundant unlabeled images that are cheap to collect, turning them into a training signal.
- Compatible with strong modern instance segmentation backbones (Mask R-CNN, SOLOv2, QueryInst).
- Consistency regularization improves robustness to photometric and geometric variations at no extra labeling cost.
- Iterative pseudo-label refinement progressively improves model quality as training proceeds.
- Pseudo-label noise can accumulate and degrade performance if confidence thresholds are set too low or the model is weak.
- Performance depends heavily on the domain match between labeled and unlabeled images.
- Training is more complex than fully supervised learning, requiring careful tuning of confidence thresholds, loss weights, and augmentation strategies.
- Evaluation requires a held-out labeled validation set that must not be used for pseudo-label generation.
Frequently asked
How many labeled images are typically needed?
Practical results vary by domain, but many studies show meaningful gains over fully supervised baselines with as few as 1–10% of images labeled. For COCO-style benchmarks, even 1% labeled data (roughly 1,200 images) can yield substantial gains when unlabeled data is abundant.
How do I set the confidence threshold for pseudo-labels?
A common starting range is 0.7–0.9 for box scores and 0.5–0.7 for mask IoU. Tune on a small validation set: too low introduces noise; too high discards useful unlabeled images. Some methods use adaptive thresholds that adjust per class.
Is a teacher-student (mean teacher) setup required?
Not strictly, but the mean-teacher design — where a slowly updated copy of the model generates pseudo-labels for the faster-updating student — is widely adopted because it produces more stable pseudo-labels than using the student model directly.
How does this differ from weakly supervised instance segmentation?
Weakly supervised methods use cheaper annotation types (bounding boxes, image tags) for all images. Semi-supervised methods use full pixel-level masks for a small subset and no annotations at all for the rest. Both reduce labeling cost, but by different mechanisms.
Can I apply this to video instance segmentation?
Yes. Temporal consistency across frames serves as an additional self-supervisory signal, and several recent methods extend the pseudo-label framework to video, propagating masks across frames using optical flow or tracking modules.
Sources
- Hu, H., Wei, P., Zheng, H., Bai, X., Wei, Y., & Chen, Y. (2021). Semi-supervised Semantic Segmentation via Adaptive Equalization Learning. Advances in Neural Information Processing Systems (NeurIPS), 34, 22106–22118. link ↗
- Xu, M., Zhang, Z., Wei, F., Hu, H., Bai, X., & Jiang, Y.-G. (2021). End-to-End Semi-Supervised Object Detection with Soft Teacher. IEEE/CVF International Conference on Computer Vision (ICCV), 3060–3069. link ↗
How to cite this page
ScholarGate. (2026, June 3). Semi-supervised Instance Segmentation. ScholarGate. https://scholargate.app/en/deep-learning/semi-supervised-instance-segmentation
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Instance SegmentationDeep learning↔ compare
- Self-supervised Vision TransformerDeep learning↔ compare
- Semantic SegmentationDeep learning↔ compare
- Semi-supervised Convolutional Neural NetworkDeep learning↔ compare
- Semi-supervised Object DetectionDeep learning↔ compare
- Weakly Supervised Instance SegmentationDeep learning↔ compare