Dogs-in-the-wild

野狗数据集是用于细粒度分类任务的大规模犬类数据集。该数据集在类别覆盖率、数据量和注释质量方面超过了类似的现有数据集。
该数据集包含362种犬的299,458图像，比斯坦福犬数据集大15倍。训练集包含258,474图像，测试集包含40,984图像。
犬类列表是通过组合多个来源 (例如维基百科)，然后通过搜索引擎 (例如谷歌、百度) 抓取图像和众包来检查每个图像的标签来生成的。通过应用混淆矩阵和人工验证，清除了少于100个图像的小类别，并合并了非常相似的类别。整个标签过程进行了三次，以确保标签质量。
示例数据:
注: train.json

The Wild Dog Dataset is a large-scale canine dataset designed for fine-grained classification tasks. It outperforms similar existing datasets in terms of category coverage, data volume, and annotation quality.

This dataset contains 299,458 images across 362 canine breeds, which is 15 times larger than the Stanford Dogs Dataset. The training set consists of 258,474 images, while the test set contains 40,984 images.

The list of canine breeds was generated by combining multiple sources (e.g., Wikipedia), then scraping images via search engines (e.g., Google, Baidu) and employing crowdsourcing to verify the label of each image. Small categories with fewer than 100 images were removed, and highly similar categories were merged through the application of confusion matrices and manual verification. The entire labeling process was conducted three times to ensure the quality of the labels.

Example Data: Note: train.json

搜集汇总

背景与挑战

背景概述

Dogs-in-the-wild是一个大规模犬类细粒度分类数据集，包含362种犬的约30万张图像，数据量是斯坦福犬数据集的15倍。该数据集通过搜索引擎抓取图像并经过严格的人工验证流程确保标签质量，由百度于2018年发布。