We construct the Web-iNat5000 dataset, which contains 5000 fine-grained categories and more than 800k training images. Web-iNat5000 is the largest webly-supervised fine-grained dataset, covering various meta-classes (e.g., plants, insecta, birds, reptiles, fungi, protozoa, mollusks, animals, etc.).
For evaluation, we construct a human-labeled evaluation dataset. This evaluation set contains 150k images (30 images per category).
We construct the iNatHash-1000 dataset which consists of 1,000 sub-categories while each sub-category contains 500 images. For each sub-category, 400 images are used as the training set and validation set, while the remaining 100 images are used to test the model. We randomly select 20 images from each sub-category as the queries, and the gallery set is formed out of all the remaining instances.
Participants are given query images and, for each query, are expected to retrieve all database images containing the same sub-category. MAP@500 is used for evaluation. Note that, participants have to provide the 48-bit hash codes for all images in both query set and gallery set.