Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
DeepEMD: Differentiable Earth Mover's Distance for Few-Shot Learning
Chi Zhang, Yujun Cai, Guosheng Lin, Chunhua Shen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use cross-reference weights instead of equal weights in EMD (ablation) | DeepEMD-FCN · miniImageNet 1-shot 5-way | accuracy: 56.95 -> 61.13 | Improved |
| use EMD instead of dense connections with cross-reference weights (ablation) | DeepEMD-FCN · miniImageNet 1-shot 5-way | accuracy: 55.41 -> 61.13 | Improved |
| set EMD weights by K-means clustering instead of equal weights (K=10) (ablation) | DeepEMD-FCN · miniImageNet 1-shot 5-way | accuracy: 56.95 -> 56.25 | Hurt |
| set EMD weights by Mean-shift with auto bandwidth (ablation) | DeepEMD-FCN · miniImageNet 1-shot 5-way | accuracy: 56.95 -> 53.56 | Hurt |
| replace global-vector metric with EMD on local features | DeepEMD-FCN · miniImageNet 1-shot 5-way | accuracy: 63.08 -> 65.91 | Improved |
| add feature pyramid P_feat(5,2,1) (ablation) | DeepEMD-FCN · miniImageNet 1-shot 5-way | accuracy: 65.91 -> 66.50 | Improved |
| add image pyramid P_grid(5,3,2) (ablation) | DeepEMD-Grid · miniImageNet 1-shot 5-way | accuracy: 67.31 -> 67.83 | Improved |
| use a finer 6x6 grid instead of 3x3 (ablation) | DeepEMD-Grid · miniImageNet 1-shot 5-way | accuracy: 67.31 -> 63.61 | Hurt |
| sample more patches (9 to 25) (ablation) | DeepEMD-Sampling · miniImageNet 1-shot 5-way | accuracy: 68.09 -> 68.77 | Improved |
| add self-supervised rotation prediction in pre-training (ablation) | DeepEMD-Sampling · miniImageNet 1-shot 5-way | accuracy: 68.77 -> 69.17 | Improved |
| apply random rotation as data augmentation only (ablation) | DeepEMD-Sampling · miniImageNet 1-shot 5-way | accuracy: 68.77 -> 67.98 | Hurt |
| use OpenCV simplex solver instead of QPTH interior point at inference | DeepEMD-FCN · EMD layer time 5x5 features 256 dim | time: 8.595 s -> 0.018 s | Improved |
Closest papers by meaning
- DeepEMD: Few-Shot Image Classification With Differentiable Earth Mover's Distance and Structured Classifiers (CVPR 2020 · similarity 0.99)
- EMD Metric Learning (AAAI 2018 · similarity 0.88)
- Revisiting Local Descriptor Based Image-To-Class Measure for Few-Shot Learning (CVPR 2019 · similarity 0.87)
- Joint Distribution Matters: Deep Brownian Distance Covariance for Few-Shot Classification (CVPR 2022 · similarity 0.87)
- A Baseline for Few-Shot Image Classification (ICLR 2020 · similarity 0.86)
- UNEM: UNrolled Generalized EM for Transductive Few-Shot Learning (CVPR 2025 · 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.
