Research map / Detection and segmentation / Open-set and few-shot detection
CGMGM: A Cross-Gaussian Mixture Generative Model 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 CGMGM, a few-shot segmentation method that models the joint pixel-category distribution of support and query features with cross-image Gaussian mixture models fit by EM, and feeds the Bayesian posterior to a parametric decoder.
Problem
Few-shot semantic segmentation must extract enough guidance from a few annotated support images; prototype and cosine-similarity methods fail to model the joint distribution of pixel and category, leading to incomplete foreground and background mis-segmentation.
Approach
Fit separate foreground and background Gaussian mixture models over features from both support and query images (using an initial query mask from a double-branch matching module refined by cycle-consistency), then compute per-pixel Bayesian posteriors of foreground as guidance for a parametric ASPP-based decoder.
Key results
- On PASCAL-5i with ResNet-50, CGMGM reaches 69.85 mIoU in 1-shot (vs. BAM 67.81) and 72.09 in 5-shot (vs. 70.91).
- On COCO-20i with ResNet-50, CGMGM improves over BAM by 1.17 mIoU in 1-shot and 0.85 in 5-shot (47.40 and 52.01 mean mIoU).
- Ablation on PASCAL-5i 1-shot: adding IMG, IMR and C&Q to the BAM baseline progressively reaches 69.85 mIoU; C&Q alone drops performance by 0.86.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace BAM with CGMGM | FSS with VGG-16 · PASCAL-5i 1-shot | mIoU (%): 64.41 -> 66.04 | Helped |
| Replace BAM with CGMGM | FSS with VGG-16 · PASCAL-5i 5-shot | mIoU (%): 68.76 -> 69.65 | Helped |
| Replace BAM with CGMGM | FSS with ResNet-50 · PASCAL-5i 1-shot | mIoU (%): 67.81 -> 69.85 | Helped |
| Replace BAM with CGMGM | FSS with ResNet-50 · PASCAL-5i 5-shot | mIoU (%): 70.91 -> 72.09 | Helped |
| Replace BAM with CGMGM | FSS with ResNet-50 · COCO-20i 1-shot | mIoU (%): 46.23 -> 47.40 | Helped |
| Replace BAM with CGMGM | FSS with ResNet-50 · COCO-20i 5-shot | mIoU (%): 51.16 -> 52.01 | Helped |
| Replace BAM with CGMGM | FSS with ResNet-50 · PASCAL-5i 1-shot | FB-IoU: 79.71 -> 80.51 | Helped |
| Replace BAM with CGMGM | FSS with ResNet-50 · PASCAL-5i 5-shot | FB-IoU: 82.18 -> 83.05 | Helped |
| Add CDGMG&QCPE alone (ablation) | BAM baseline on PASCAL-5i 1-shot · PASCAL-5i 1-shot | mIoU (%): 67.81 -> 66.95 | Hurt |
| Add Initial Mask Generating to CDGMG&QCPE (ablation) | BAM baseline on PASCAL-5i 1-shot · PASCAL-5i 1-shot | mIoU (%): 66.95 -> 68.27 | Helped |
| Add Initial Mask Refining to CDGMG&QCPE (ablation) | BAM baseline on PASCAL-5i 1-shot · PASCAL-5i 1-shot | mIoU (%): 66.95 -> 68.58 | Helped |
| Add Initial Mask Generating, Initial Mask Refining, and CDGMG&QCPE (ablation) | BAM baseline on PASCAL-5i 1-shot · PASCAL-5i 1-shot | mIoU (%): 67.81 -> 69.85 | Helped |
Surprising findings
- Using only the GMM posterior and decoder (C&Q) without the initial mask module decreased mIoU below the BAM baseline (66.95 vs 67.81).
Limitations
- Relies on BAM initialization and base learner; the gain over BAM is modest (about 2 mIoU on PASCAL-5i 1-shot).
- Ablation shows the GMM module alone hurts performance without the initial mask generation/refinement.
Experimental setup
Datasets: PASCAL-5i COCO-20i
Benchmarks: PASCAL-5i COCO-20i
Metrics: mIoU FB-IoU
Baselines: PFENet HSNet DPCN NTRENet BAM ASGNet CyCTR DCAMA IPMT PFNet
Builds on: BAM PFENet CyCTR GMM-Seg
Closest papers by meaning
- A Surprisingly Simple Approach to Generalized Few-Shot Semantic Segmentation (NeurIPS 2024 · similarity 0.88)
- Foreground-Covering Prototype Generation and Matching for SAM-Aided Few-Shot Segmentation (AAAI 2025 · similarity 0.87)
- Label-Efficient Few-Shot Semantic Segmentation with Unsupervised Meta-Training (AAAI 2024 · similarity 0.87)
- DSV-LFS: Unifying LLM-Driven Semantic Cues with Visual Features for Robust Few-Shot Segmentation (CVPR 2025 · similarity 0.87)
- Unleashing the Potential of the Diffusion Model in Few-shot Semantic Segmentation (NeurIPS 2024 · similarity 0.86)
- Unlocking the Potential of Pre-trained Vision Transformers for Few-Shot Semantic Segmentation through Relationship Descriptors (CVPR 2024 · similarity 0.86)
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.