There are 2 different ImageNet datasets: ImageNet 1k usually referred to in papers as just ImageNet and the full ImageNet dataset also called ImageNet 22k. Is the ImageNet 1k a subset of the 22k? And if so are the 1k categories represented the same way in the 22k version or maybe images of a certain class in the 1k version are aggregated from several different classes in the 22k version (like different dog breeds are merged into the "dog" class)? Can I somehow cut the 1k version from the 22k version and say the rest of the 22k version except the 1k version is a completely different dataset or even an out of distribution dataset for the 1k version? Any links to some trusted sources would be cool. Thanks
Yes, the full ImageNet is a superset of the 1k version (see Benchmarking Neural Network Robustness to Common Corruptions and Perturbations). The 1k synsets (a.k.a. classes) corresponding to the 1k version are also present in the full 22k version. So there are 21841 classes in total, 1000 of them are exactly the same as in the 1k version (and images are exactly the same), and the other 20841 can be treated as OOD to the 1k version dataset (at least they are treated so in one of the papers by Hendrycks et al, Using Self-Supervised Learning Can Improve Model Robustness and Uncertainty.