Research map / Detection and segmentation / Open-set and few-shot detection
Relevant Intrinsic Feature Enhancement Network for 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 RiFeNet, a few-shot segmentation network with a training-only unlabeled branch that enforces intra-class feature consistency plus multi-level (global support + local query) prototype generation and interaction, built on CyCTR.
Problem
Few-shot semantic segmentation suffers from semantic ambiguity of foreground instances across intra-class variation and inter-class similarity that confuses pixel-level foreground-background classification.
Approach
An unlabeled branch (sharing query-branch parameters, trained with weak/strong augmentation pseudo-labels and a dice loss) regularizes foreground consistency during training only, and multi-granularity prototypes (global support prototype plus local query prototypes from PFENet-style prior masks) are interacted to widen inter-class separation.
Key results
- On PASCAL-5i 1-shot with ResNet-50, RiFeNet reaches 67.1 mIoU vs. 64.6 for DCAMA and 64.0 for CyCTR baseline.
- On COCO 1-shot with ResNet-50, RiFeNet reaches 44.1 mIoU vs. 43.3 for DCAMA.
- Ablation on PASCAL-5i 1-shot: unlabeled branch alone 66.1, multi-level prototypes alone 66.2, both 67.1 (baseline 64.0).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the unlabeled branch (ablation) | CyCTR on PASCAL-5ᶦ · 1-shot, Res-50 | mIoU: 64.0 -> 66.1 | Helped |
| add multi-level prototypes (ablation) | CyCTR on PASCAL-5ᶦ · 1-shot, Res-50 | mIoU: 64.0 -> 66.2 | Helped |
| add both the unlabeled branch and multi-level prototypes (ablation) | CyCTR on PASCAL-5ᶦ · 1-shot, Res-50 | mIoU: 64.0 -> 67.1 | Helped |
| replace local query prototypes with global query prototypes (ablation) | RiFeNet on PASCAL-5ᶦ · 1-shot, Res-50 | mIoU: 67.1 -> 66.3 | Hurt |
| remove channel-wise attention from local prototypes (ablation) | RiFeNet on PASCAL-5ᶦ · 1-shot, Res-50 | mIoU: 66.8 -> 67.1 | Hurt |
| remove query local prototype guidance from the unlabeled branch (ablation) | RiFeNet on PASCAL-5ᶦ · 1-shot, Res-50 | mIoU: 65.8 -> 67.1 | Hurt |
| double training iterations without adding the unlabeled branch (ablation) | CyCTR on PASCAL-5ᶦ · 1-shot, Res-50 | mIoU: 66.2 -> 66.0 | No clear effect |
| increase unlabeled images from 0 to 1 per task (ablation) | RiFeNet on PASCAL-5ᶦ · 1-shot, Res-50 | mIoU: 66.2 -> 66.6 | Helped |
| increase unlabeled images from 1 to 2 per task (ablation) | RiFeNet on PASCAL-5ᶦ · 1-shot, Res-50 | mIoU: 66.6 -> 67.1 | Helped |
| increase unlabeled images from 2 to 3 per task (ablation) | RiFeNet on PASCAL-5ᶦ · 1-shot, Res-50 | mIoU: 67.1 -> 66.0 | Hurt |
Surprising findings
- Using the unlabeled branch without query-prototype guidance performs worse than the baseline (65.8 vs 64.0 is better; reported as 65.8 vs 67.1 with guidance), and gains shrink in 5-shot as unlabeled-to-labeled ratio drops.
- Increasing unlabeled images beyond 2 per task decreases accuracy, attributed to feature attention shifting to the unlabeled branch.
Limitations
- Unlabeled branch gains diminish in the 5-shot setting as labeled-to-unlabeled ratio grows.
- Performance is sensitive to the number of unlabeled images; more than 2 per task hurts accuracy.
Experimental setup
Datasets: PASCAL-5i COCO
Benchmarks: PASCAL-5i COCO
Metrics: mIoU FB-IoU
Baselines: PPNet PMM PFENet CyCTR HSNet ASGNet SSP DCAMA RPMMs DAN
Builds on: CyCTR PFENet PPNet DICE loss
Closest papers by meaning
- Deconfound Semantic Shift and Incompleteness in Incremental Few-shot Semantic Segmentation (AAAI 2025 · similarity 0.86)
- Unlocking the Potential of Pre-trained Vision Transformers for Few-Shot Semantic Segmentation through Relationship Descriptors (CVPR 2024 · similarity 0.85)
- Enhancing Generalized Few-Shot Semantic Segmentation via Effective Knowledge Transfer (AAAI 2025 · similarity 0.85)
- A Surprisingly Simple Approach to Generalized Few-Shot Semantic Segmentation (NeurIPS 2024 · similarity 0.85)
- Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 2024 · similarity 0.85)
- Simple Semantic-Aided Few-Shot Learning (CVPR 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.