This demo utilizes a large convolutional neural network model trained on images from the 2012 ImageNet competition. There are 1,000 possible classes that the model learned about from 1.3 million training images. Your image is cropped to 256x256 from the center of the image and fed into the model to predict which classes of objects are present. Out of this 256x256 image, 10 crops of size 224x224 at the 4 corners, center, and their horizontal flips are used to make the model more robust. The probabilities that the model assigns to each class for the 10 crops are averaged to get the final predictions for your image.