Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Reinforcing Generated Images via Meta-Learning for One-Shot Fine-Grained Visual Recognition
Satoshi Tsutsui, Yanwei Fu, David Crandall
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 |
|---|---|---|---|
| add FinetuneGAN-generated images directly as augmentation | ProtoNet · CUB 5-way-1-shot | accuracy (%): 81.73 -> 79.40 | Hurt |
| add the learned image fusion module mixing real and generated images (ablation) | ProtoNet · CUB 5-way-1-shot | accuracy (%): 81.73 -> 84.13 | Improved |
| add the learned image fusion module mixing real and generated images (ablation) | ProtoNet · NAB 5-way-1-shot | accuracy (%): 87.91 -> 89.19 | Improved |
| add horizontal flip augmentation | MetaIRNet (+FinetuneGAN) · CUB 5-way-1-shot | accuracy (%): 84.13 -> 84.80 | Improved |
| use FinetuneGAN images instead of FreezeDGAN images for mixing (ablation) | MetaIRNet · CUB 5-way-1-shot | accuracy (%): 81.93 -> 84.13 | Improved |
| use FinetuneGAN images instead of jittered originals for mixing (ablation) | MetaIRNet · CUB 5-way-1-shot | accuracy (%): 82.69 -> 84.13 | Improved |
| add hand-designed 3x3 binary mixed images to training data | Original (nearest neighbor) · CUB validation 5-way-1-shot | accuracy (%): 70.72 -> 71.50 | Improved |
| add plain generated images to training data | Logistic regression · CUB validation 5-way-1-shot | accuracy (%): 75.39 -> 74.08 | Hurt |
| add plain generated images to training data | Softmax regression · CUB validation 5-way-1-shot | accuracy (%): 74.61 -> 73.55 | Hurt |
| add hand-designed mixed images to training data | Softmax regression · CUB validation 5-way-1-shot | accuracy (%): 74.61 -> 75.40 | Improved |
| use 5x5 mixing weights instead of 3x3 (ablation) | MetaIRNet · CUB 5-way-1-shot | accuracy (%): 84.13 -> 83.63 | Hurt |
| increase number of generated images n_aug from 1 to 5 (ablation) | MetaIRNet · CUB 5-way-1-shot | accuracy (%) | No clear effect |
Closest papers by meaning
- Meta-Reinforced Synthetic Data for One-Shot Fine-Grained Visual Recognition (NeurIPS 2019 · similarity 0.95)
- Meta Internal Learning (NeurIPS 2021 · similarity 0.88)
- MetaGAN: An Adversarial Approach to Few-Shot Learning (NeurIPS 2018 · similarity 0.87)
- CVAE-GAN: Fine-Grained Image Generation Through Asymmetric Training (ICCV 2017 · similarity 0.86)
- Fine-Grained Image-to-Image Transformation Towards Visual Recognition (CVPR 2020 · similarity 0.86)
- Generative Dual Adversarial Network for Generalized Zero-Shot Learning (CVPR 2019 · 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.
