The Dogs-in-the-Wild dataset is a large-scale dog dataset for fine-grained classification tasks. The dataset exceeds similar existing datasets in terms of category coverage, data volume, and annotation quality.

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

The dog list was generated by combining multiple sources (such as Wikipedia), and then crawling images through search engines (such as Google and Baidu) to check the labels of each image in a crowdsourcing manner. Small categories with less than 100 images were eliminated, and extremely similar categories were merged by applying confusion matrices and manual verification. The entire annotation process was performed three times to ensure the quality of the annotations.

Example data:

{
    "info": "Dogs-in-the-Wild",
    "split": "train",
    "annotations": [
        {
            "name": "image/train/3472791206,1521450563.jpg",
            "image id": 0,
            "category id": 312
        }

        ...
    ]
}

Note: train.json