Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Cross-Layer and Cross-Sample Feature Optimization Network for Few-Shot Fine-Grained Image Classification
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 C2-Net, which refines features across the last two backbone layers via cross-layer correlation reconstruction (CLFR) and aligns query to support prototypes via channel reweighting and learned position warping (CSFA) for few-shot fine-grained image classification.
Problem
Few-shot fine-grained image classification suffers from over-fitting due to sample-level noise introduced by multi-layer fine-grained features, and from feature mismatch between support and query samples caused by spatial position and angle variation.
Approach
Cross-layer correlation matrices reconstruct and denoise features from the last two backbone layers, and a cross-sample module adjusts query features via trainable channel activation and a learned position-matching transform toward the support prototype.
Key results
- On CUB (Conv-4, raw images), C2-Net reaches 78.66% 1-shot and 89.43% 5-shot, versus BiFRN's 74.36% and 88.64%.
- On Stanford Dogs (ResNet-12), C2-Net achieves 75.50% 1-shot vs BiFRN's 72.54%; on Stanford Cars (ResNet-12) it reports 88.96% 1-shot but 95.16% 5-shot, below BiFRN's 96.34% 5-shot.
- On meta-iNat (Conv-4), C2-Net reports 71.47% 1-shot and 85.47% 5-shot, above the best listed baseline MCL (64.66% / 81.31%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use C2-Net instead of BiFRN | Conv-4 few-shot classifier · Stanford-Dogs, 1-shot | accuracy (%): 61.39 ± 0.23 -> 66.42 ± 0.50 | Helped |
| Use C2-Net instead of BiFRN | Conv-4 few-shot classifier · Stanford-Dogs, 5-shot | accuracy (%): 78.86 ± 0.15 -> 81.23 ± 0.34 | Helped |
| Use C2-Net instead of AGPF | Conv-4 few-shot classifier · Stanford-Cars, 1-shot | accuracy (%): 78.14 ± 0.84 -> 81.29 ± 0.45 | Helped |
| Use C2-Net instead of BiFRN | Conv-4 few-shot classifier · Stanford-Cars, 5-shot | accuracy (%): 90.66 ± 0.11 -> 91.08 ± 0.26 | Helped |
| Use C2-Net instead of BiFRN | ResNet-12 few-shot classifier · Stanford-Dogs, 1-shot | accuracy (%): 72.54 ± 0.22 -> 75.50 ± 0.49 | Helped |
| Use C2-Net instead of BiFRN | ResNet-12 few-shot classifier · Stanford-Dogs, 5-shot | accuracy (%): 85.86 ± 0.13 -> 87.65 ± 0.28 | Helped |
| Use C2-Net instead of BiFRN | ResNet-12 few-shot classifier · Stanford-Cars, 1-shot | accuracy (%): 88.43 ± 0.17 -> 88.96 ± 0.37 | Helped |
| Use C2-Net instead of BiFRN | ResNet-12 few-shot classifier · Stanford-Cars, 5-shot | accuracy (%): 96.34 ± 0.07 -> 95.16 ± 0.20 | Hurt |
| Use C2-Net instead of BiFRN | Conv-4 few-shot classifier · CUB (raw images), 1-shot | accuracy (%): 74.36 ± 0.20 -> 78.66 ± 0.46 | Helped |
| Use C2-Net instead of PaCL | Conv-4 few-shot classifier · CUB (raw images), 5-shot | accuracy (%): 88.75 ± 0.38 -> 89.43 ± 0.28 | Helped |
| Add CLFR to the baseline (ablation) | Conv-4 few-shot classifier · CUB (raw images), 1-shot | accuracy (%): 54.87 -> 69.81 | Helped |
| Add CSFA to the baseline (ablation) | Conv-4 few-shot classifier · CUB (raw images), 1-shot | accuracy (%): 54.87 -> 76.51 | Helped |
Surprising findings
- On Stanford Cars with ResNet-12, C2-Net's 5-shot accuracy (95.16%) is lower than BiFRN's (96.34%), so it is not best in every setting.
Limitations
- Gains are smaller in some settings, e.g., Stanford Cars 5-shot with ResNet-12 trails BiFRN.
- Only the last two layers are used (L=2), chosen for performance and complexity reasons; deeper multi-layer fusion is not explored.
Experimental setup
Datasets: Stanford Dogs Stanford Cars CUB-200-2011 meta-iNat tiered meta-iNat
Benchmarks: Stanford Dogs Stanford Cars CUB-200-2011 meta-iNat tiered meta-iNat
Metrics: 5-way 1-shot accuracy 5-way 5-shot accuracy
Baselines: DN4 CovaMNet MattML ATL-Net BSNet LRPABN TOAN OLSA DAN AGPF PaCL HelixFormer
Builds on: ProtoNet FRN OLSA AGPF BiFRN
Closest papers by meaning
- Few-Shot Fine-Grained Image Classification with Progressively Feature Refinement and Continuous Relationship Modeling (AAAI 2025 · similarity 0.90)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.87)
- Twofold Debiasing Enhances Fine-Grained Learning with Coarse Labels (AAAI 2025 · similarity 0.86)
- Discriminative Sample-Guided and Parameter-Efficient Feature Space Adaptation for Cross-Domain Few-Shot Learning (CVPR 2024 · similarity 0.86)
- Strong Baselines for Parameter-Efficient Few-Shot Fine-Tuning (AAAI 2024 · similarity 0.85)
- Adapt Before Comparison: A New Perspective on Cross-Domain Few-Shot Segmentation (CVPR 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.