Research map / Label-efficient and robust learning / Knowledge distillation
Mutual Learning for SAM Adaptation: A Dual Collaborative Network Framework for Source-Free Domain Transfer
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.
Proposes a collaborative mutual learning framework that dynamically assigns teacher/student roles between two LoRA-tuned SAM image encoders for source-free SAM adaptation.
Problem
Source-free adaptation of the Segment Anything Model to target domains with large distribution shifts, where fixed teacher-student self-training does not guarantee the teacher outperforms the student.
Approach
Two LoRA-adapted SAM encoders act as peers; each iteration, the one whose foreground features are more similar to a frozen reference SAM becomes teacher, combined with direct alignment, reverse distillation, and a triplet relationship loss.
Key results
- Average mIoU of 71.50 on COCO and 76.45 on Pascal VOC, versus 69.53 for WeSAM on COCO.
- Average mIoU of 70.74 on CAMO and 64.66 on COD10K, improving over WeSAM (67.29 and 62.78).
- Average mIoU of 83.29 on Kvasir-SEG and 84.02 on OCID, the best among compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use collaborative mutual learning instead of WeSAM adaptation | SAM · COCO | average mIoU: 69.53 -> 71.50 | Helped |
| use collaborative mutual learning instead of WeSAM adaptation | SAM · Pascal VOC | average mIoU: 73.71 -> 76.45 | Helped |
| use collaborative mutual learning instead of WeSAM adaptation | SAM · CAMO | average mIoU: 67.29 -> 70.74 | Helped |
| use collaborative mutual learning instead of WeSAM adaptation | SAM · COD10K | average mIoU: 62.78 -> 64.66 | Helped |
| use collaborative mutual learning instead of WeSAM adaptation | SAM · kvasir-SEG | average mIoU: 76.03 -> 83.29 | Helped |
| use collaborative mutual learning instead of WeSAM adaptation | SAM · OCID | average mIoU: 81.88 -> 84.02 | Helped |
| add reverse distillation to direct alignment (ablation) | collaborative SAM adaptation · CAMO | average mIoU: 68.83 -> 70.22 | Helped |
| add reverse distillation to direct alignment (ablation) | collaborative SAM adaptation · COD10K | average mIoU: 63.21 -> 64.13 | Helped |
| add triplet relationship loss to direct alignment (ablation) | collaborative SAM adaptation · CAMO | average mIoU: 68.83 -> 69.75 | Helped |
| add triplet relationship loss to direct alignment (ablation) | collaborative SAM adaptation · COD10K | average mIoU: 63.21 -> 63.75 | Helped |
| add both reverse distillation and triplet relationship losses to direct alignment (ablation) | collaborative SAM adaptation · CAMO | average mIoU: 68.83 -> 70.74 | Helped |
| add both reverse distillation and triplet relationship losses to direct alignment (ablation) | collaborative SAM adaptation · COD10K | average mIoU: 63.21 -> 64.66 | Helped |
Surprising findings
- The student network can surpass the fixed teacher during training under severe domain shift, motivating dynamic roles.
Limitations
- Relies on a frozen reference SAM and foreground-similarity reliability metric whose robustness is not analyzed across domains.
- Hyperparameters (lambda values, margin, threshold) chosen without broad sensitivity analysis reported.
Experimental setup
Datasets: COCO Pascal VOC Kvasir-SEG CAMO COD10K OCID
Benchmarks: COCO Pascal VOC CAMO COD10K Kvasir-SEG OCID
Metrics: mIoU average mIoU GAIN
Baselines: WeSAM WDASS SHOT TENT TRIBE DePT soft Teacher
Builds on: WeSAM LoRA Segment Anything Model deep mutual learning SHOT
Closest papers by meaning
- Multi-Source Domain Adaptation With Collaborative Learning for Semantic Segmentation (CVPR 2021 · similarity 0.88)
- InfoSAM: Fine-Tuning the Segment Anything Model from An Information-Theoretic Perspective (ICML 2025 · similarity 0.86)
- Semantic-aware SAM for Point-Prompted Instance Segmentation (CVPR 2024 · similarity 0.86)
- Every SAM Drop Counts: Embracing Semantic Priors for Multi-Modality Image Fusion and Beyond (CVPR 2025 · similarity 0.86)
- Improving the Generalization of Segmentation Foundation Model under Distribution Shift via Weakly Supervised Adaptation (CVPR 2024 · similarity 0.85)
- SAM-REF: Introducing Image-Prompt Synergy during Interaction for Detail Enhancement in the Segment Anything Model (CVPR 2025 · 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.