Research map / Detection and segmentation / Weakly supervised detection and segmentation
Scribbles for All: Benchmarking Scribble Supervised Segmentation Across Datasets
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Scribbles for All, an algorithm that automatically generates scribble labels from dense segmentation masks, releases four scribble datasets (s4Pascal, s4Cityscapes, s4KITTI360, s4ADE20K), and benchmarks SASS methods on them.
Problem
Scribble-supervised semantic segmentation lacks challenging benchmark datasets: PascalVOC is the only popular scribble dataset and is too easy, and manual scribble annotation is costly.
Approach
An automatic scribble generator that takes dense semantic masks, applies size-dependent erosion, and fits polynomial curves through randomly sampled edge point pairs and the noisy center of mass, with validity checks that prevent crossing class boundaries; it yields scribbles statistically similar to human ScribbleSup labels.
Key results
- On s4Pascal, SASFormer reaches 79.0 mIoU vs 79.5 on ScribbleSup, and EMA reaches 78.8 vs 79.2, showing synthetic scribbles are near-equivalent to human scribbles.
- Relative performance to full supervision drops to roughly 80% on s4Cityscapes, s4ADE20K and s4KITTI360 (e.g., TEL 64.5 mIoU on s4Cityscapes; EMA 67.7 mIoU; SASFormer 41.4 mIoU on s4ADE20K), versus about 90-92% on Pascal.
- TEL drops by less than 2% mIoU when scribble length is reduced to a tenth, while SASFormer is most sensitive to shorter scribbles.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use s4Pascal synthetic scribbles instead of ScribbleSup hand-drawn scribbles | EMA · PascalVOC | mIoU: 79.2 -> 78.8 | Hurt |
| Use s4Pascal synthetic scribbles instead of ScribbleSup hand-drawn scribbles | SASformer · PascalVOC | mIoU: 79.5 -> 79.0 | Hurt |
| Use s4Pascal synthetic scribbles instead of ScribbleSup hand-drawn scribbles | TEL · PascalVOC | mIoU: 76.4 -> 76.8 | Helped |
| Use s4Pascal synthetic scribbles instead of ScribbleSup hand-drawn scribbles | AGMM · PascalVOC | mIoU: 76.8 -> 73.8 | Hurt |
| Shorten scribbles to one tenth of their original length (ablation) | TEL · scribble-length ablation | mIoU: less than 2 % mIoU | Hurt |
| Shorten scribbles (ablation) | SASformer · scribble-length ablation | mIoU | Hurt |
| Shorten scribbles (ablation) | EMA · scribble-length ablation | mIoU | Hurt |
Surprising findings
- Scribble-supervised methods on Pascal reach about 90-92% of fully supervised performance, but the gap widens substantially on more complex datasets, with the naive mean-teacher baseline competitive on some datasets.
- Some SASS methods (AGMM-SASS) are more sensitive to the label distribution change, dropping about 3% mIoU on s4Pascal relative to ScribbleSup.
Limitations
- Requires an existing dense semantic segmentation ground truth as input, so it cannot be applied to new domains without dense labels.
- Scribbles from the generator produce more scribbles per image than human labels on Pascal (5.11 vs 3.89), and the generator cannot draw over do-not-care regions.
- Base datasets have varying licenses (e.g., Cityscapes proprietary), so users must obtain the base datasets separately.
Experimental setup
Datasets: PascalVOC 2012 ScribbleSup Cityscapes KITTI360 ADE20K
Benchmarks: PascalVOC 2012 Cityscapes KITTI360 ADE20K
Metrics: mIoU relative performance vs fully supervised share of labeled pixels boundary distance statistics
Baselines: ScribbleSup TEL AGMM-SASS SASFormer EMA mean teacher fully supervised
Builds on: ScribbleSup Tree Energy Loss AGMM-SASS SASFormer SegFormer
Closest papers by meaning
- Scribble Hides Class: Promoting Scribble-Based Weakly-Supervised Semantic Segmentation with Its Class Label (AAAI 2024 · similarity 0.93)
- Scribble-Supervised Semantic Segmentation with Prototype-based Feature Augmentation (ICML 2024 · similarity 0.92)
- Exploring Simple Open-Vocabulary Semantic Segmentation (CVPR 2025 · similarity 0.86)
- From SAM to CAMs: Exploring Segment Anything Model for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.86)
- W2P: Switching from Weak Supervision to Partial Supervision for Semantic Segmentation (AAAI 2024 · similarity 0.86)
- Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.