Research map / Label-efficient and robust learning / Knowledge distillation
EdgeSAM: Prompt-In-the-Loop Distillation for SAM
Chong Zhou, Xiangtai Li, Chen Change Loy, Bo Dai
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 prompt-in-the-loop distillation to encoder-only KD (ablation) | EdgeSAM (RepViT) · SA-1K center point | mIoU: 64.6 -> 67.5 | Improved |
| add prompt-in-the-loop distillation to encoder-only KD (ablation) | EdgeSAM (RepViT) · SA-1K GT box | mIoU: 82.0 -> 83.0 | Improved |
| add prompt-in-the-loop KD to TinyViT backbone (ablation) | TinyViT-5M · SA-1K center point | mIoU: 63.7 -> 68.0 | Improved |
| add prompt-in-the-loop KD to EfficientViT backbone (ablation) | EfficientViT-B1 · SA-1K center point | mIoU: 60.9 -> 65.8 | Improved |
| add granularity-prior RPN (ablation) | EdgeSAM · COCO center point | mIoU: 48.0 -> 54.3 | Improved |
| add granularity-prior RPN (ablation) | EdgeSAM · 2080 Ti speed | FPS: 164.3 -> 123.9 | Hurt |
| use RepViT-M1 backbone instead of TinyViT-5M (FPN alignment) (ablation) | EdgeSAM encoder-only KD · SA-1K | FPS: 114.2 -> 164.3 | Improved |
| two-stage training instead of joint training (ablation) | EdgeSAM · SA-1K center point | mIoU: 63.2 -> 67.4 | Improved |
| supervise decoder with mask loss only instead of image feature loss (ablation) | EdgeSAM · SA-1K GT box | mIoU: 71.6 -> 83.1 | Improved |
| add attention-map or IoU alignment loss to mask loss (ablation) | EdgeSAM decoder · SA-1K GT box | mIoU: 83.1 -> 82.8 (attn), 82.5 (IoU) | No clear effect |
| add one prompt sampling loop (ablation) | EdgeSAM · SA-1K center +2 pt. | mIoU: 72.4 -> 79.0 | Improved |
| add a second prompt sampling loop (ablation) | EdgeSAM · SA-1K center +2 pt. | mIoU: 79.0 -> 79.2 | No clear effect |
Closest papers by meaning
- EdgeTAM: On-Device Track Anything Model (CVPR 2025 · similarity 0.89)
- BA-SAM: Scalable Bias-Mode Attention Mask for Segment Anything Model (CVPR 2024 · similarity 0.85)
- Stable Segment Anything Model (ICLR 2025 · similarity 0.84)
- TinySAM: Pushing the Envelope for Efficient Segment Anything Model (AAAI 2025 · similarity 0.84)
- EfficientSAM: Leveraged Masked Image Pretraining for Efficient Segment Anything (CVPR 2024 · similarity 0.83)
- SAM-I2V: Upgrading SAM to Support Promptable Video Segmentation with Less than 0.2% Training Cost (CVPR 2025 · similarity 0.83)
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.
