Research map / Label-efficient and robust learning / Semi-supervised learning
S4M: Boosting Semi-Supervised Instance Segmentation with SAM
Heeji Yoon, Heeseong Shin, Eunbeen Hong, Hyunwook Choi, Hansang Cho, Daun Jeong, Seungryong Kim
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add structural distillation loss from SAM to the teacher (ablation) | Mask2Former teacher · Cityscapes 5% labeled | mask AP: 14.9 -> 17.2 | Improved |
| add structural distillation loss from SAM to the teacher (ablation) | Mask2Former teacher · Cityscapes 10% labeled | mask AP: 19.0 -> 22.5 | Improved |
| add structural distillation loss from SAM to the teacher (ablation) | Mask2Former teacher · COCO 10% labeled | mask AP: 30.1 -> 30.3 | No clear effect |
| apply structural distillation to teacher only (ablation) | S4M · Cityscapes 10% labeled | mask AP: 30.1 -> 32.8 | Improved |
| also apply structural distillation to the student (ablation) | S4M with teacher SD · Cityscapes 10% labeled | mask AP: 32.8 -> 29.4 | Hurt |
| distill SAM decoder features instead of encoder features (ablation) | teacher structural distillation · Cityscapes 10% labeled | mask AP: 21.4 -> 22.4 | Improved |
| use structural loss instead of feature distillation (ablation) | teacher distillation from SAM encoder · Cityscapes 10% labeled | mask AP: 21.1 -> 21.4 | Improved |
| add pseudo-label refinement with SAM (ablation) | S4M with SD · Cityscapes 10% labeled | mask AP: 29.2 -> 32.3 | Improved |
| add augmentation with refined pseudo-labels (ARP) alone (ablation) | baseline student · Cityscapes 10% labeled | mask AP: 28.4 -> 27.8 | Hurt |
| add ARP on top of SD and pseudo-label refinement (ablation) | S4M · Cityscapes 10% labeled | mask AP: 32.3 -> 32.8 | Improved |
| use K sampled points instead of a single point prompt for SAM refinement (ablation) | S4M · Cityscapes 10% labeled | mask AP: 30.4 -> 32.8 | Improved |
| use K sampled points instead of mask prompt for SAM refinement (ablation) | S4M · Cityscapes 10% labeled | mask AP: 24.3 -> 32.8 | Improved |
Closest papers by meaning
- Semantic-aware SAM for Point-Prompted Instance Segmentation (CVPR 2024 · similarity 0.90)
- Instance-Specific and Model-Adaptive Supervision for Semi-Supervised Semantic Segmentation (CVPR 2023 · similarity 0.90)
- WISH: Weakly Supervised Instance Segmentation using Heterogeneous Labels (CVPR 2025 · similarity 0.89)
- SAM-CP: Marrying SAM with Composable Prompts for Versatile Segmentation (ICLR 2025 · similarity 0.88)
- From SAM to CAMs: Exploring Segment Anything Model for Weakly Supervised Semantic Segmentation (CVPR 2024 · similarity 0.88)
- SAM as the Guide: Mastering Pseudo-Label Refinement in Semi-Supervised Referring Expression Segmentation (ICML 2024 · similarity 0.88)
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.
