MSCOCO-GLT
收藏资源简介：
对于MSCOCO-GLT，尽管我们可以直接从MSCOCO-attribute获得属性注释，但每个对象可能具有多个属性，这使得严格平衡属性分布变得令人望而却步，因为每次我们对具有稀有属性的对象进行采样时，它通常也包含共同出现的频繁属性。因此，如上图1所示，我们收集具有相对更平衡的属性分布 (通过最小化属性的STD) 的图像子集，以用作属性均衡测试集。其他长尾拆分和类拆分的构造与ImageNet-GLT相同。 请注意，Train-CBL拆分的大小明显小于Train-GLT，因为一个超级头目类 “人” 包含超过40% 的整个数据，使得按类数据重新平衡非常昂贵。Train-CBL上较差的结果进一步证明了长尾分类在实际应用中的重要性，因为大型长尾数据集比小型平衡数据集更好。
For MSCOCO-GLT, while we can directly obtain attribute annotations from MSCOCO-attribute, each object may possess multiple attributes, which makes it daunting to strictly balance the attribute distribution: whenever we sample objects with rare attributes, they typically also contain frequently co-occurring common attributes. Therefore, as shown in Figure 1 above, we collected an image subset with a relatively more balanced attribute distribution (by minimizing the standard deviation (STD) of attributes) to serve as the attribute-balanced test set. The construction of other long-tail splits and class splits is identical to that of ImageNet-GLT. Note that the size of the Train-CBL split is significantly smaller than that of Train-GLT, as the super-majority class "person" accounts for over 40% of the entire dataset, making class-wise data rebalancing extremely costly. Poorer performance on Train-CBL further demonstrates the importance of long-tail classification in real-world applications, as large-scale long-tail datasets outperform small-scale balanced ones.
