salman53/waterbirds

Waterbirds数据集是通过从Caltech-UCSD Birds-200-2011 (CUB)数据集的照片中裁剪出鸟类，并将其转移到Places数据集的背景上构建的。该数据集使用CUB的官方训练-测试分割，并随机选择20%的训练数据作为验证集。训练集中，95%的水鸟被放置在水背景上，其余5%在陆背景上；95%的陆鸟被放置在陆背景上，其余5%在水背景上。开发/测试集中，50%的水鸟在水背景上，其余50%在陆背景上；50%的陆鸟在陆背景上，其余50%在水背景上。这种构建方式导致训练集和开发/测试集之间存在分布偏移。数据集旨在研究分布鲁棒性，特别是在群体偏移情况下的泛化能力。

The Waterbirds dataset is constructed by cropping out birds from photos in the Caltech-UCSD Birds-200-2011 (CUB) dataset and transferring them onto backgrounds from the Places dataset. The official train-test split of the CUB dataset is used, with 20% of the training data randomly chosen to serve as a validation set. In the training set, 95% of waterbirds are placed against water backgrounds and the remaining 5% against land backgrounds, while 95% of landbirds are against land backgrounds with the remaining 5% against water. In the dev/test set, 50% of waterbirds are against water backgrounds and the remaining 50% against land backgrounds, and 50% of landbirds are against land backgrounds with the remaining 50% against water. This construction creates a distribution shift between train and dev/test, making it suitable for studying distributional robustness and generalization under group shifts.