imagenet

ImageNet is the large-scale human-annotated image database that has shaped modern image classification more than any other single resource. The full database organises over fourteen million images according to the WordNet noun hierarchy, but in practice 'ImageNet' almost always refers to the subset used in the ImageNet Large Scale Visual Recognition Challenge (ILSVRC): roughly 1.28 million training images spanning 1000 mutually exclusive object classes, with 50000 validation and 100000 test images.

Its importance is historical as well as practical. The 2012 winner, AlexNet, cut top-5 error dramatically using a deep convolutional network trained on GPUs, and that result is widely credited with igniting the modern deep-learning era in vision. Since then ImageNet top-1 accuracy has been the de facto leaderboard on which architectures such as VGG, ResNet, EfficientNet, and Vision Transformers, together with their training recipes, are compared.

Beyond benchmarking, ImageNet is the default pretraining corpus: weights from a network trained to classify ImageNet-1k are reused as initialisation for countless downstream tasks. Be aware of its documented limitations, including label noise, some ambiguous or offensive categories in the larger hierarchy, and a geographic and cultural bias in the photographs, since these can propagate into any model pretrained on it.