摘要
本文解决了学习比训练标签所提供的表示更为精细的问题。这使得在仅用粗略标签注释的图像集合中实现细粒度类别检索成为可能。我们的网络通过最近邻分类器目标进行学习，并采用了一种受自监督学习启发的实例损失。通过共同利用粗略标签和潜在的细粒度隐空间，该方法显著提高了类别级检索方法的准确性。我们的策略在检索或分类图像时，其细粒度超过了训练时可用的细粒度，优于所有竞争方法。此外，它还提高了向细粒度数据集迁移学习任务的准确性，从而在五个公开基准测试（如iNaturalist-2018）上确立了新的最先进水平。
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| fine-grained-image-classification-on-food-101 | Grafit (RegNet-8GF) | Accuracy: 93.7 |
| fine-grained-image-classification-on-oxford | Grafit (RegNet-8GF) | Accuracy: 99.1% |
| fine-grained-image-classification-on-stanford | Grafit (RegNet-8GF) | Accuracy: 94.7% |
| image-classification-on-cifar-100 | Grafit (ResNet-50) | Percentage correct: 83.7 |
| image-classification-on-flowers-102 | Grafit (RegNet-8GF) | Accuracy: 99.1% |
| image-classification-on-imagenet | Grafit (ResNet-50) | Hardware Burden: Operations per network pass: Top 1 Accuracy: 79.6% |
| image-classification-on-inaturalist-2018 | ResNet-50 | Top-1 Accuracy: 69.8% |
| image-classification-on-inaturalist-2018 | RegNet-8GF | Top-1 Accuracy: 81.2% |
| image-classification-on-inaturalist-2019 | Grafit (RegnetY 8GF) | Top-1 Accuracy: 84.1 |
| learning-with-coarse-labels-on-cifar100 | Grafit | Recall@1: 60.57 Recall@10: 89.21 Recall@2: 71.13 Recall@5: 82.32 |
| learning-with-coarse-labels-on-imagenet32 | Grafit | Recall@1: 18.13 Recall@10: 46.64 Recall@2: 25.46 Recall@5: 37.19 |
| learning-with-coarse-labels-on-stanford | Grafit | Recall@1: 74.02 Recall@10: 87.91 Recall@2: 78.82 Recall@5: 84.13 |
| learning-with-coarse-labels-on-stanford-cars | Grafit | Recall@1: 42.30 Recall@10: 81.74 Recall@2: 54.79 Recall@5: 71.1 |