Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Few-Shot Fine-Grained Image Classification with Progressively Feature Refinement and Continuous Relationship Modeling
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 SUITED, which adds top-down progressive feature refinement across backbone layers (PFRM) and a GCN-style continuous one-to-many prototype-query relationship module (CRMM) for few-shot fine-grained image classification.
Problem
Few-shot fine-grained image classification needs more discriminative class prototypes from the backbone and more accurate modeling of query-prototype similarity than discrete one-to-one pairwise matching provides.
Approach
A top-down feedback mechanism that uses higher-layer features to gate lower-layer features to form multi-scale prototypes, plus a GCN-based continuous one-to-many prototype-to-query relationship module that refines query features during training only.
Key results
- With ResNet-12, SUITED reaches 86.02% (1-shot) and 94.13% (5-shot) on CUB, exceeding C2-Net (83.31 / 92.18).
- On meta-iNat with the reported setting, SUITED scores 74.72% 1-shot and 87.44% 5-shot, versus C2-Net's 71.47 / 85.47.
- Ablation on Conv-4 CUB: PFRM alone gives 77.54 / 88.27 and the full model 79.73 / 90.05 (1-shot / 5-shot).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add SUITED instead of C2-Net | FS-FGIC · CUB-200-2011, Conv-4 1-shot | accuracy: 78.66 -> 79.73 | Helped |
| Add SUITED instead of C2-Net | FS-FGIC · meta-iNat 1-shot | accuracy: 71.47 -> 74.72 | Helped |
| Add PFRM to the backbone (ablation) | SUITED · CUB-200-2011, Conv-4 1-shot | accuracy: 68.87 -> 77.54 | Helped |
| Add PFRM to the backbone (ablation) | SUITED · Stanford-Cars, Conv-4 1-shot | accuracy: 67.08 -> 78.34 | Helped |
| Add CRMM to the backbone (ablation) | SUITED · CUB-200-2011, Conv-4 1-shot | accuracy: 68.87 -> 73.98 | Helped |
| Add CRMM to the backbone (ablation) | SUITED · Stanford-Cars, Conv-4 1-shot | accuracy: 67.08 -> 73.27 | Helped |
| Use both PFRM and CRMM instead of PFRM alone (ablation) | SUITED · CUB-200-2011, Conv-4 1-shot | accuracy: 77.54 -> 79.73 | Helped |
| Use both PFRM and CRMM instead of PFRM alone (ablation) | SUITED · Stanford-Cars, Conv-4 1-shot | accuracy: 78.34 -> 82.21 | Helped |
| Model continuous relationships from prototypes to queries (P to Q) (ablation) | CRMM · CUB-200-2011, Conv-4 1-shot | accuracy: 78.61 -> 79.73 | Helped |
| Model continuous relationships from prototypes to queries (P to Q) (ablation) | CRMM · Stanford-Cars, Conv-4 1-shot | accuracy: 81.29 -> 82.21 | Helped |
Surprising findings
- Removing CRMM at test time has little impact on performance, so it acts as a training-only regularizer.
Limitations
- Gains over the strongest prior method are modest on some settings (e.g., ResNet-12 CUB 1-shot vs MLI 85.94 is not exceeded, and Stanford Cars results are close to prior methods).
- Compute cost and training time of the GCN-based relationship module are not analyzed.
Experimental setup
Datasets: CUB-200-2011 Stanford Dogs Stanford Cars meta-iNat tiered meta-iNat
Benchmarks: CUB-200-2011 Stanford Dogs Stanford Cars meta-iNat tiered meta-iNat
Metrics: accuracy (5-way 1-shot, 5-way 5-shot, 95% CI)
Baselines: ProtoNet BSNet PoseNorm FEAT FRN OLSA DAN TDM AGPF PaCL BiFRN C2-Net
Builds on: ProtoNet GCN FRN BiFRN C2-Net
Closest papers by meaning
- Cross-Layer and Cross-Sample Feature Optimization Network for Few-Shot Fine-Grained Image Classification (AAAI 2024 · similarity 0.90)
- Fine-grained Classes and How to Find Them (ICML 2024 · similarity 0.85)
- Advancing Fine-Grained Classification by Structure and Subject Preserving Augmentation (NeurIPS 2024 · similarity 0.85)
- Twofold Debiasing Enhances Fine-Grained Learning with Coarse Labels (AAAI 2025 · similarity 0.85)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.84)
- Strong Baselines for Parameter-Efficient Few-Shot Fine-Tuning (AAAI 2024 · similarity 0.84)
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.