摘要
计算机视觉任务，如图像分类、图像检索和少样本学习，目前主要由欧几里得嵌入和球面嵌入主导，最终关于类别归属或相似度程度的决策是通过线性超平面、欧几里得距离或球面测地距离（余弦相似度）来做出的。在本研究中，我们展示了在许多实际场景中，双曲嵌入提供了一个更好的替代方案。
代码仓库
KhrulkovV/hyperbolic-image-embeddings
官方
pytorch
GitHub 中提及
nalexai/hyperlib
tf
GitHub 中提及
leymir/hyperbolic-image-embeddings
官方
pytorch
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| few-shot-image-classification-on-cub-200-5 | Hyperbolic ProtoNet | Accuracy: 72.22 |
| few-shot-image-classification-on-cub-200-5-1 | Hyperbolic ProtoNet | Accuracy: 60.52 |
| few-shot-image-classification-on-mini-2 | Hyperbolic ProtoNet | Accuracy: 51.57 |
| few-shot-image-classification-on-mini-3 | Hyperbolic ProtoNet | Accuracy: 66.27 |
| few-shot-image-classification-on-omniglot-1-1 | Hyperbolic ProtoNet | Accuracy: 95.9% |
| few-shot-image-classification-on-omniglot-1-2 | Hyperbolic ProtoNet | Accuracy: 99.0 |
| few-shot-image-classification-on-omniglot-5-1 | Hyperbolic ProtoNet | Accuracy: 98.15% |
| few-shot-image-classification-on-omniglot-5-2 | Hyperbolic ProtoNet | Accuracy: 99.4 |