BAR

该数据集包含了手动挑选的真实世界动作与背景配对图像，其中背景与动作具有相关性。此外，该数据集存在类别不平衡的问题，某些类别的图像数量明显较少。数据规模为1641张训练图像、300张验证图像和654张测试图像，分为6个目标类别。该数据集的任务是视觉偏见缓解。

This dataset contains manually curated real-world paired images of actions and their correlated backgrounds. Furthermore, this dataset suffers from class imbalance, with certain target categories having significantly fewer image samples. It comprises 1,641 training images, 300 validation images, and 654 test images, which are divided into 6 target categories. The task of this dataset is visual bias mitigation.