Meta-Dataset
收藏资源简介:
元数据集是一个大型的很少的学习基准。此数据集不会对小样本任务施加限制 (不需要固定的方法和镜头),因此它代表了更真实的场景。 数据集由来自不同域的10个数据集组成: ILSVRC-2012 (由涉及1,000个类别的自然图像组成的ImageNet数据集。)Omniglot (手写字符,包含1,623类别) 飞机 (飞机图像数据集,包含100类别) CUB-200-2011 (鸟数据集,包含200类别) 可描述纹理 (不同种类的纹理图像,包含43个类别) 快速绘制 (涵盖345个不同类别的黑白草图) 真菌 (涵盖1,500个类别的大型蘑菇数据集) VGG花 (涵盖类别的102花图像数据集),交通标志 (德国交通标志图像,包含43个类别) MSCOCO (图片收集自Flickr,包含80个类别) 元数据集中的交通标志 (GTSRB) 和COCO数据集不参与培训,仅用于验证或测试。其余8个数据集根据70%/15% 的比例大致分为训练/验证/测试集。
This meta-dataset is a large-scale few-shot learning benchmark. This dataset imposes no restrictions on few-shot tasks (no fixed methodologies or shot counts), thus representing a more realistic research scenario. This benchmark consists of 10 datasets spanning diverse domains: 1. ILSVRC-2012: the ImageNet dataset composed of natural images across 1,000 categories 2. Omniglot: a handwritten character dataset containing 1,623 categories 3. Aircraft: an aircraft image dataset with 100 categories 4. CUB-200-2011: a bird image dataset with 200 categories 5. Describable Textures: a dataset of diverse texture images with 43 categories 6. Quick Draw: a black-and-white sketch dataset covering 345 distinct categories 7. Fungi: a large-scale mushroom dataset encompassing 1,500 categories 8. VGG Flowers: a 102-category flower image dataset 9. Traffic Signs: a German traffic sign image dataset with 43 categories 10. MSCOCO: an image collection sourced from Flickr, with 80 categories The Traffic Signs (GTSRB) and MSCOCO datasets within this meta-dataset are not used for training, and are only utilized for validation or testing. The remaining 8 datasets are roughly split into training/validation/test subsets at a ratio of 70%/15%/15% respectively.

- Meta-Dataset首次发表于ICML(国际机器学习大会),由Google Research团队提出,旨在解决元学习中的数据多样性问题。
- Meta-Dataset被广泛应用于多个研究项目中,特别是在小样本学习领域,显著提升了模型的泛化能力。
- Meta-Dataset的扩展版本发布,增加了更多的数据源和类别,进一步丰富了数据集的多样性。
- Meta-Dataset在多个国际竞赛中被采用,成为评估元学习算法性能的标准数据集之一。
- 1Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few ExamplesGoogle Research · 2019年
- 2Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few ExamplesGoogle Research · 2020年
- 3Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few ExamplesGoogle Research · 2021年
- 4Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few ExamplesGoogle Research · 2022年
- 5Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few ExamplesGoogle Research · 2023年



