摘要
现实世界中的数据大多不平衡且呈长尾分布，但深度模型在存在频繁类别的情况下难以识别稀有类别。通常，类别可以伴随诸如文本描述的辅助信息，但如何利用这些信息进行不平衡长尾数据的学习尚不完全明确。此类描述主要被用于（广义）零样本学习（ZSL），这表明带有类别描述的ZSL也可能对长尾分布有益。我们介绍了一种名为DRAGON的后期融合架构，用于带有类别描述的长尾学习。该架构能够（1）逐样本纠正偏向头部类别的偏差；以及（2）融合类别描述中的信息以提高尾部类别的准确性。我们还引入了新的基准数据集CUB-LT、SUN-LT和AWA-LT，这些数据集基于现有的属性学习数据集，并包含带有类别描述的Imagenet-LT版本。DRAGON在新基准上超越了现有最先进模型的表现，同时也在现有的GFSL带类别描述（GFSL-d）和标准（仅视觉）长尾学习基准ImageNet-LT、CIFAR-10、CIFAR-100和Places365上取得了新的最佳性能。
代码仓库
dvirsamuel/DRAGON
官方
tf
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| generalized-few-shot-learning-on-awa2 | DRAGON | Per-Class Accuracy (1-shot): 67.1 Per-Class Accuracy (10-shots): 81.9 Per-Class Accuracy (2-shots): 69.1 Per-Class Accuracy (20-shots): 83.3 Per-Class Accuracy (5-shots): 76.7 |
| generalized-few-shot-learning-on-sun | DRAGON | Per-Class Accuracy (1-shot): 41.0 Per-Class Accuracy (10-shots): 48.2 Per-Class Accuracy (2-shots): 43.8 Per-Class Accuracy (5-shots): 46.7 |
| long-tail-learning-on-cifar-10-lt-r-10 | smDRAGON | Error Rate: 11.84 |
| long-tail-learning-on-cifar-10-lt-r-100 | smDRAGON | Error Rate: 20.37 |
| long-tail-learning-on-cifar-100-lt-r-10 | smDRAGON | Error Rate: 41.23 |
| long-tail-learning-on-cifar-100-lt-r-100 | smDRAGON | Error Rate: 56.50 |
| long-tail-learning-on-imagenet-lt | smDRAGON | Top-1 Accuracy: 42.0 |
| long-tail-learning-on-places-lt | smDRAGON | Top-1 Accuracy: 38.1 |
| long-tail-learning-with-class-descriptors-on | DRAGON + Bal'Loss | Long-Tailed Accuracy: 66.5 Per-Class Accuracy: 60.1 |
| long-tail-learning-with-class-descriptors-on | DRAGON | Long-Tailed Accuracy: 67.7 Per-Class Accuracy: 57.8 |
| long-tail-learning-with-class-descriptors-on-1 | DRAGON + Bal'Loss | Long-Tailed Accuracy: 38.5 Per-Class Accuracy: 36.1 |
| long-tail-learning-with-class-descriptors-on-1 | DRAGON | Long-Tailed Accuracy: 40.4 Per-Class Accuracy: 34.8 |
| long-tail-learning-with-class-descriptors-on-2 | DRAGON + Bal'Loss | Long-Tailed Accuracy: 92.2 Per-Class Accuracy: 76.2 |
| long-tail-learning-with-class-descriptors-on-2 | DRAGON | Long-Tailed Accuracy: 94.1 Per-Class Accuracy: 74.1 |
| long-tail-learning-with-class-descriptors-on-3 | DRAGON + Bal'Loss | Per-Class Accuracy: 53.5 |
| long-tail-learning-with-class-descriptors-on-3 | DRAGON | Per-Class Accuracy: 51.2 |