ImageNet-4K consists of 4,214 classes and 1,050 images per class. The dataset is split into 1,000 images for the database and 50 images for the query set. The ImageNet-4K classes are a subset of ImageNet-21K classes which do not overlap with ImageNet-1K and contain 1,050 or more images each. The database/query sets contain contain ∼4.2M and ∼200K total images respectively. The list of images curated together to construct ImageNet-4K can be found here. The mapping from ImageNet ID to class name and definition can be found in IDtoClass.txt.
To construct ImageNet-4K first obtain ImageNet-21K (the full ImageNet from Deng et al 2009). Once obtained create a subset of the data using the list of images here. Please email Aditya Kusupati for further questions and assistance regarding the dataset.