Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
FeatWalk: Enhancing Few-Shot Classification through Local View Leveraging
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 FeatWalk, a plug-in module that fuses a global-view feature with weighted random local-view features via prototype-similarity weights, training only a classifier head on top of a frozen pre-trained feature extractor for few-shot classification.
Problem
Pre-trained or meta-trained feature extractors for few-shot learning may neglect local features that are crucial for downstream tasks, since they learn mainly image-level representations.
Approach
Sample random local patches as extra views, weight each local view by its cosine similarity to a per-class prototype, and fuse them with the global feature by a walk from the global feature toward local features; only a classifier head after the fusion is trained on a frozen backbone.
Key results
- On miniImageNet 5-way 1-shot, FeatWalk reaches 70.21% vs. 69.14% for FGM and 69.71% for RENet-ventral (ResNet-12).
- On CUB with ResNet-18, FeatWalk reaches 85.67% (1-shot) and 95.44% (5-shot), improving over DeepBDC by 1.66% and 1.42%.
- Cross-domain (miniImageNet to CUB) 5-shot accuracy is 83.60% vs. 80.16% for DeepBDC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add FeatWalk with weighted local views to DeepBDC | DeepBDC · miniImageNet and tieredImageNet 5-way 1-shot and 5-shot; CUB 5-way 1-shot and 5-shot | accuracy: +0.50%; +1.37%; +0.96%; +0.92%; +1.66%; +1.42% | Helped |
| add FeatWalk to DeepBDC for cross-domain evaluation | DeepBDC · CUB cross-domain 5-shot | accuracy: 83.60% vs. 80.16% | Helped |
| add local-view sampling and FeatWalk rather than the single-branch FC baseline (ablation) | DeepBDC · miniImageNet and CUB 5-way 1-shot and 5-shot | accuracy: 67.80 ± 0.45 -> 70.21 ± 0.44; 85.21 ± 0.28 -> 87.38 ± 0.27; 83.97 ± 0.39 -> 85.67 ± 0.38; 94.40 ± 0.17 -> 95.44 ± 0.16 | Helped |
| apply data augmentation to the single-branch classifier instead of FeatWalk (ablation) | DeepBDC single-branch FC classifier · miniImageNet and CUB 5-way 1-shot and 5-shot | accuracy: 67.80 ± 0.45 -> 68.11 ± 0.43; 85.21 ± 0.28 -> 85.04 ± 0.28; 83.97 ± 0.39 -> 83.04 ± 0.40; 94.40 ± 0.17 -> 93.68 ± 0.18 | No clear effect |
| use equal importance for all local views instead of learned weights (ablation) | FeatWalk · miniImageNet and CUB 5-way 1-shot and 5-shot | accuracy: 69.66 ± 0.45 -> 70.21 ± 0.44; 86.70 ± 0.27 -> 87.38 ± 0.27; 84.91 ± 0.38 -> 85.67 ± 0.38; 94.92 ± 0.17 -> 95.44 ± 0.16 | Hurt |
| replace FC with LR classifier and cosine similarity with Euclidean distance (ablation) | FeatWalk · miniImageNet and CUB 5-way 1-shot and 5-shot | accuracy: 69.78±0.44 vs. 70.21±0.44; 86.96±0.27 vs. 87.38±0.27; 84.81±0.40 vs. 85.67±0.38; 94.61±0.18 vs. 95.44±0.16 | Hurt |
| replace FC classifier with LR (ablation) | FeatWalk with cosine similarity · miniImageNet and CUB 5-way 1-shot and 5-shot | accuracy: 70.10±0.44 vs. 70.21±0.44; 87.32±0.27 vs. 87.38±0.27; 85.43±0.39 vs. 85.67±0.38; 95.20±0.17 vs. 95.44±0.16 | Mixed |
| replace cosine similarity with Euclidean distance (ablation) | FeatWalk with FC classifier · miniImageNet and CUB 5-way 1-shot and 5-shot | accuracy: 70.09±0.44 vs. 70.21±0.44; 87.21±0.27 vs. 87.38±0.27; 85.84±0.38 vs. 85.67±0.38; 95.51±0.16 vs. 95.44±0.16 | Mixed |
Surprising findings
- Simple data augmentation with the classifier head gave no improvement over the FC baseline, while equal-weight local fusion helped, indicating learned local view weighting matters.
Limitations
- Evaluation limited to image classification on three benchmarks; applicability to medical images and text is untested.
- Relies on a strong pre-trained backbone that is frozen; only the classifier head is trained.
Experimental setup
Datasets: miniImageNet tieredImageNet CUB
Benchmarks: miniImageNet tieredImageNet CUB
Metrics: 5-way 1-shot accuracy 5-way 5-shot accuracy
Baselines: DN4 BML MCL DeepEMD v2 CovNet Baseline++ ProtoNet Good-Embed DeepBDC FGM RENet-ventral FEAT
Builds on: DeepBDC Good-Embed ProtoNet
Closest papers by meaning
- Few-Shot Learning via Embedding Adaptation With Set-to-Set Functions (CVPR 2020 · similarity 0.88)
- CAD: Co-Adapting Discriminative Features for Improved Few-Shot Classification (CVPR 2022 · similarity 0.87)
- Pushing the Limit of Fine-Tuning for Few-Shot Learning: Where Feature Reusing Meets Cross-Scale Attention (AAAI 2024 · similarity 0.87)
- Finding Task-Relevant Features for Few-Shot Learning by Category Traversal (CVPR 2019 · similarity 0.87)
- Making Large Vision Language Models to Be Good Few-Shot Learners (AAAI 2025 · similarity 0.86)
- AMU-Tuning: Effective Logit Bias for CLIP-based 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.