Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
APSeg: Auto-Prompt Network for Cross-Domain Few-Shot Semantic Segmentation
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 APSeg, which adapts frozen SAM (ViT-base) to cross-domain few-shot segmentation using cycle-consistency-based dual prototype anchor transformation and a meta prompt generator that replaces manual prompts.
Problem
Cross-domain few-shot semantic segmentation (CD-FSS): segmenting unseen classes in target domains (medical, remote sensing, etc.) with few labeled samples after training only on natural-image PASCAL VOC, where existing FSS methods and SAM both degrade.
Approach
Augments support prototypes with pseudo query foreground/background prototypes selected by cycle-consistent matching to compute a domain-agnostic transformation matrix (DPAT), and generates SAM prompt embeddings automatically via a meta-learned prompt generator (MPG) instead of manual points/boxes.
Key results
- APSeg reaches 61.30% average 1-shot mIoU and 65.09% 5-shot across four CD-FSS datasets, versus 56.06%/61.99% for PATNet.
- Improves Chest X-ray 1-shot mIoU to 84.10% from PATNet's 66.61% (+17.49%).
- Ablation: adding MPG and DPAT to the baseline raises 1-shot average from 42.86% to 61.30%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use APSeg instead of PATNet | CD-FSS · Average 1-shot | mIoU: 56.06 -> 61.30 | Helped |
| Use APSeg instead of PATNet | CD-FSS · Average 5-shot | mIoU: 61.99 -> 65.09 | Helped |
| Add MPG to the baseline (ablation) | APSeg · ISIC 1-shot | mIoU: 44.55 -> 36.67 | Hurt |
| Add MPG to the baseline (ablation) | APSeg · Average 1-shot | mIoU: 42.86 -> 45.84 | Helped |
| Add DPAT to Baseline + MPG (ablation) | APSeg · Average 1-shot | mIoU: 45.84 -> 61.30 | Helped |
| Use both sparse and dense embeddings instead of sparse embeddings only (ablation) | MPG · ISIC 1-shot | mIoU: 41.52 -> 45.43 | Helped |
| Use both sparse and dense embeddings instead of dense embeddings only (ablation) | MPG · ISIC 1-shot | mIoU: 44.09 -> 45.43 | Helped |
| Set output feature channels to 64 instead of 128 (ablation) | APSeg · Average 1-shot | mIoU: 60.03 -> 61.30 | Helped |
| Set output feature channels to 64 instead of 256 (ablation) | APSeg · Average 1-shot | mIoU: 58.65 -> 61.30 | Helped |
| Generate 4 sparse embeddings instead of 1 (ablation) | MPG · Average 1-shot | mIoU: 60.75 -> 61.30 | Helped |
| Generate 8 sparse embeddings instead of 4 (ablation) | MPG · Average 1-shot | mIoU: 61.30 -> 60.65 | Hurt |
Surprising findings
- PerSAM (manual visual prompts on SAM) reaches only 37.56% average 1-shot mIoU, below several CNN-based FSS methods.
Limitations
- Performance on Deepglobe (35.94% 1-shot) and FSS-1000 trails some baselines such as HSNet and PATNet-ViT.
- Relies on frozen SAM features; SAM's weak zero-shot performance on distant domains still bounds results.
Experimental setup
Datasets: PASCAL VOC 2012 SBD Chest X-ray ISIC FSS-1000 Deepglobe
Benchmarks: Chest X-ray ISIC FSS-1000 Deepglobe
Metrics: mIoU
Baselines: AMP PGNet PANet CaNet RPMMs PFENet RePRI HSNet PerSAM PATNet RestNet
Builds on: Segment Anything Model PATNet PerSAM HSNet PFENet
Closest papers by meaning
- Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 2024 · similarity 0.90)
- SAM-Aware Graph Prompt Reasoning Network for Cross-Domain Few-Shot Segmentation (AAAI 2025 · similarity 0.90)
- Bridge the Points: Graph-based Few-shot Segment Anything Semantically (NeurIPS 2024 · similarity 0.89)
- Adapter Naturally Serves as Decoupler for Cross-Domain Few-Shot Semantic Segmentation (ICML 2025 · similarity 0.89)
- Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 2024 · similarity 0.89)
- Domain-Rectifying Adapter for Cross-Domain Few-Shot Segmentation (CVPR 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.