Weakly Supervised Semantic Segmentation
Weakly Supervised Semantic Segmentation (WSSS) · Also known as: WSSS, weak-label segmentation, image-level supervised segmentation, weakly-labeled pixel classification
Weakly Supervised Semantic Segmentation (WSSS) trains pixel-level scene parsers using only cheap, coarse annotations — typically image-level class tags — instead of costly dense pixel masks. By generating proxy pseudo-labels from a classification network (via Class Activation Maps or similar localisation cues) and iteratively refining them, WSSS brings full-supervision accuracy within reach at a fraction of the 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
Choose WSSS when pixel-level annotation is prohibitively expensive — medical imaging, remote sensing, industrial inspection — and image-level or bounding-box labels are available. It is well-suited for large image databases where exhaustive labelling is impractical. Do not use it when precise boundary accuracy is critical (e.g., autonomous driving safety systems), when very few labeled images are available (few-shot is a better fit), or when fully annotated data already exists, as full supervision will always outperform weak supervision given equal training data.
Strengths & limitations
- Dramatically reduces annotation cost: image-level tags are orders of magnitude cheaper than pixel masks.
- Leverages large unlabeled or weakly-labeled web datasets that would be unusable for fully supervised methods.
- Produces full-resolution pixel predictions at inference — no coarse labels needed at test time.
- Compatible with modern backbones (ViT, Swin Transformer) that naturally produce richer localisation maps.
- Active research area with rapidly closing performance gap to fully supervised counterparts.
- Performance still lags behind fully supervised segmentation, especially at object boundaries.
- CAM-derived seeds are biased toward discriminative parts (e.g., the head of a bird) and miss other object regions.
- Multi-stage pipelines add complexity: CAM generation, affinity learning, CRF refinement, and final training are distinct steps.
- Heavily class-imbalanced datasets and small or thin objects are poorly localised by CAMs.
Frequently asked
What is the most common weak supervision signal used in WSSS?
Image-level class tags are by far the most popular because they are cheapest to obtain and often already available in image databases. Bounding boxes and scribbles provide stronger localisation cues but require more annotation effort.
How close is WSSS to fully supervised segmentation in accuracy?
On benchmarks such as PASCAL VOC 2012, top WSSS methods reach 75–78 mIoU with image-level supervision, compared to around 80–85 mIoU for fully supervised methods. The gap narrows further with modern transformer backbones.
Is a CRF mandatory in the WSSS pipeline?
Dense CRF post-processing is common because it sharpens pseudo-mask boundaries, but it is not mandatory. Many recent works replace it with learned boundary modules or self-attention mechanisms that achieve comparable or better refinement.
Can WSSS be combined with a small number of fully annotated images?
Yes. Combining a few fully labeled images with many weakly labeled ones (a semi-weakly supervised setting) typically boosts performance substantially and is a practical choice when a small annotation budget exists.
Which backbone works best for seed generation?
Vision Transformers (ViT, DeiT) produce spatially smoother and more complete CAMs than CNNs because their global self-attention captures full object extent rather than only discriminative parts. Hybrid CNN-Transformer backbones are also popular.
Sources
- Zhou, B., Khosla, A., Lapedriza, A., Oliva, A., & Torralba, A. (2016). Learning Deep Features for Discriminative Localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 2921–2929. DOI: 10.1109/CVPR.2016.319 ↗
- Ahn, J., & Kwak, S. (2018). Learning Pixel-Wise Semantic Affinity with Image-Level Supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pp. 4109–4118. link ↗
How to cite this page
ScholarGate. (2026, June 3). Weakly Supervised Semantic Segmentation (WSSS). ScholarGate. https://scholargate.app/en/deep-learning/weakly-supervised-semantic-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.
- Object DetectionDeep learning↔ compare
- Self-supervised LearningMachine learning↔ compare
- Semantic SegmentationDeep learning↔ compare
- Semi-supervised LearningMachine learning↔ compare