AwA2 (Animals with Attributes 2)

该数据集提供了一个基准转移学习算法的平台，特别是属性基础分类和零射击学习 [1]。由于具有相同的类结构和几乎相同的特征，因此它可以替代具有属性 (AwA) 数据集 [2,3] 的原始动物。
它由50个动物类的37322个图像组成，每个图像具有预先提取的特征表示。这些类与Osherson的经典类/属性矩阵 [3,4] 对齐，从而为每个类提供85个数字属性值。使用共享属性，可以在不同的类之间传输信息。
图像数据是从公共来源 (如Flickr，2016年) 收集的。在此过程中，我们确保仅包含免费使用和重新分发许可的图像，请参阅各个许可文件的存档。如果数据集包含您拥有版权且未经免费许可的图像，请通过以下方式与我们联系:

This dataset functions as a benchmark platform for transfer learning algorithms, specifically attribute-based classification and zero-shot learning [1]. Boasting identical class structures and nearly identical feature spaces, it can serve as a substitute for the original Animals with Attributes (AwA) dataset [2,3]. It consists of 37,322 images spanning 50 animal classes, with each image accompanied by pre-extracted feature representations. These classes align with Osherson's classic class/attribute matrix [3,4], which provides 85 numerical attribute values for every class. By utilizing shared attributes, information can be transferred between distinct classes. The image data was collected from public sources (e.g., Flickr, 2016). During the collection process, we ensured that only images with licenses allowing free use and redistribution were included; please refer to the archive of individual license documents for further information. If the dataset contains images for which you hold copyright and for which no free license was granted, please contact us via the following methods:

搜集汇总

背景与挑战

背景概述

AwA2是一个用于转移学习算法基准测试的图像数据集，特别适用于属性基础分类和零射击学习研究。它包含50个动物类的37322张图像，每张图像都配有预先提取的特征和85个数字属性值，便于在类间传输信息。数据集图像来自Flickr等公共来源，并确保使用免费许可，由阿姆斯特丹大学和马克斯普朗克信息学研究所于2020年发布。