arubique/waterbirds
收藏资源简介:
Waterbirds (OCCAM布局) 数据集是一个用于图像分类和鲁棒性研究的基准数据集,特别关注虚假相关性和群体偏移。该数据集基于原始Waterbirds基准,由Sagawa等人引入,用于群体分布鲁棒优化。数据集包含鸟类图像(分为陆地鸟和水鸟两类),并与陆地或水背景合成,以创建虚假相关性(例如,陆地鸟通常出现在陆地背景上)。数据集以OCCAM代码库的布局重新组织,包含12个子场景,每个子场景对应不同的鸟类型和背景组合(如landbird_on_land、waterbird_on_water等),并提供原始图像、仅前景裁剪和仅背景裁剪版本。这些版本用于评估模型在子群体偏移、仅前景或仅背景条件下的性能。数据集用于研究深度学习模型的鲁棒性,特别是在处理虚假相关性时的泛化能力。每个子场景文件夹包含二进制类别标签(0代表水鸟,1代表陆地鸟),便于图像分类任务。数据集还包括元数据文件(metadata.csv),用于映射图像文件路径和背景信息。该数据集适用于机器学习和计算机视觉社区,用于实验和分析模型在分布偏移下的行为。
The Waterbirds (OCCAM Layout) dataset is a benchmark dataset for image classification and robustness research, with a particular focus on spurious correlations and distribution shift. It is based on the original Waterbirds benchmark introduced by Sagawa et al. for distributionally robust optimization. The dataset contains bird images categorized into two classes: land birds and water birds, which are synthesized with land or water backgrounds to create spurious correlations (e.g., land birds typically appear on land backgrounds). Reorganized according to the layout of the OCCAM codebase, the dataset includes 12 sub-scenarios, each corresponding to a unique combination of bird type and background (e.g., landbird_on_land, waterbird_on_water). It provides three versions of each image: the original full image, a foreground-only cropped version, and a background-only cropped version. These variants are used to evaluate model performance under subgroup shift, as well as foreground-only or background-only conditions. This dataset is designed to study the robustness of deep learning models, especially their generalization ability when handling spurious correlations. Each sub-scenario folder contains binary category labels, where 0 denotes water birds and 1 denotes land birds, to facilitate image classification tasks. Additionally, the dataset includes a metadata file (metadata.csv) that maps image file paths to their corresponding background information. This dataset is suitable for the machine learning and computer vision communities, supporting experiments and analyses of model behavior under distribution shift.




