ImageNet

ImageNet is a database of more than 14 million images, each labeled with one of 21,841 noun concepts from the WordNet lexical hierarchy, built by Fei-Fei Li's group at Princeton and Stanford and introduced in 2009. In everyday use, "ImageNet" usually means ImageNet-1k, the 1,000-class dataset of the 2012 ImageNet Large Scale Visual Recognition Challenge (ILSVRC), with 1,281,167 training images, 50,000 validation images, and 100,000 test images. ImageNet-1k is the standard benchmark dataset for image classification, and networks trained on it are a common starting point for transfer learning in computer vision.

How ImageNet Is Organized#

Each WordNet concept, called a synset, groups the words that name it and falls under more general concepts, so golden retriever falls under dog, which falls under animal. The ImageNet team gathered candidate images for each synset through web image searches and had Amazon Mechanical Turk workers verify every one, aiming for 500 to 1,000 clean images per concept. ImageNet-1k keeps 1,000 synsets, mostly fine-grained leaves of the hierarchy rather than broad categories: 118 of its classes are dog breeds, so a model must tell a Siberian Husky from an Alaskan Malamute as well as a school bus from a goldfinch. The full release with all 21,841 synsets is called ImageNet-21k, or ImageNet-22k, and is used mainly for large-scale pretraining.

The ILSVRC Challenge and AlexNet#

ILSVRC ran every year from 2010 to 2017 and ranked classification entries by top-5 error, the share of test images whose true class is missing from a model's five highest-scoring guesses. Top-1 accuracy, which counts only the single best guess, became the more common figure as models improved. In 2012, AlexNet, a convolutional neural network (CNN) with 60 million parameters trained on two GPUs, won with a top-5 error of 15.3% against 26.2% for the runner-up, which relied on hand-engineered features. That margin turned computer vision research toward deep learning. Later winners kept cutting the error, and in 2015 an ensemble of residual networks (ResNet) reached 3.57%, below the 5.1% top-5 error measured for a trained human annotator in the ILSVRC paper.

Pretraining and Transfer Learning#

A network trained on ImageNet-1k learns general visual features in its early layers, such as edges, textures, and simple shapes, and more object-specific features in its later layers. Reusing those weights as a backbone and fine-tuning them on a smaller dataset converges faster and reaches higher accuracy than training from random weights, with the largest gains when labeled images are few. YOLO26 classification models are trained on ImageNet-1k at 224 pixels and reach top-1 validation accuracies from 71.4% for YOLO26n-cls to 79.9% for YOLO26x-cls, and training one on a custom dataset replaces its 1,000-class output layer with one sized for the new classes.

from ultralytics import YOLO

# Load a YOLO26 nano classifier trained on the 1,000 ImageNet-1k classes
model = YOLO("yolo26n-cls.pt")

# Classify an image and select the result for that image
results = model("https://ultralytics.com/images/bus.jpg")
result = results[0]

# Print the five most likely ImageNet classes with their probabilities
for index, probability in zip(result.probs.top5, result.probs.top5conf.tolist()):
    print(f"{result.names[index]}: {probability:.3f}")

The model scores all 1,000 classes, and the loop prints the five highest, the same top-5 list that the ILSVRC error metric checks against the true label.

The largest vision models now pretrain on much bigger and noisier collections, such as the 400 million image-text pairs behind CLIP, yet ImageNet-1k remains a standard test of how good their learned features are, through fine-tuning, zero-shot classification, or linear probing, which trains only a linear classifier on frozen features.

Limitations and Follow-Up Datasets#

• Label errors: At least 6% of ImageNet-1k validation labels are wrong, and many images contain several objects but carry a single label. The ReaL labels re-annotated the validation set with every valid class per image.
• ImageNetV2: A new test set collected by repeating the original process. Models lost 11% to 14% accuracy on it while keeping nearly the same ranking, a drop the authors attribute to slightly harder images rather than years of reuse of the original test set.
• Privacy and fairness: In 2019 the ImageNet team filtered 2,702 synsets from the person subtree of the full dataset, and in 2021 it released a face-blurred version of the ILSVRC data that costs only marginal accuracy. The 1,000 ILSVRC classes were unaffected by the person-subtree filtering.
• Access terms: ImageNet does not own the copyright of its images and grants access for non-commercial research and education only. The ImageNet-1k training and validation images total about 144 GB.

ImageNet vs. Related Datasets#

• COCO: Boxes, masks, and keypoints for 80 object categories with many labeled objects per image, the standard benchmark for object detection and segmentation.
• CIFAR-10: 60,000 images of 32×32 pixels in 10 classes, small enough for experiments that train in minutes on one GPU.