ILSVRC2012

ILSVRC2012, short for the ImageNet Large Scale Visual Recognition Challenge 2012, was the second edition of the annual competition focused on large-scale image classification and object recognition. The core task remained classifying images into 1000 predefined object categories drawn from the ImageNet dataset. The event used a standardized test protocol with a held-out test set and a public validation set for benchmarking, with the top-5 error rate as a primary metric.

The competition is notable for the entry by Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton, known as AlexNet. This deep convolutional neural network consists of eight layers, including five convolutional layers followed by three fully connected layers. It employed Rectified Linear Units as activation functions, max-pooling, and local response normalization, along with data augmentation and dropout in the fully connected layers. Training was performed on two GPUs, and the model contains roughly 60 million parameters. On the ImageNet validation set, AlexNet achieved a top-5 error rate of 15.3%, a dramatic improvement over prior approaches and widely regarded as a turning point for deep learning in computer vision.

Dataset and evaluation: The ILSVRC2012 dataset comprises approximately 1.2 million training images, 50,000 validation images, and 100,000 test images across 1000 categories. Performance on the validation set was the primary benchmark, with results later reflected on the leaderboard for the competition’s test phase. The success of AlexNet helped catalyze widespread adoption of deep learning methods for visual recognition and influenced subsequent architectural developments in the field.