The CIFAR-10 and CIFAR-100 datasets are small (compared to modern standards) image datasets collected by Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. These datasets are widely used by the research community for image classification tasks. They are considered challenging, because the image quality is very low and the objects in the images are sometimes partially visible. At the same time, the datasets are convenient due to this small image size, so researchers can quickly produce results on them. CIFAR-100 increases the challenge since there are only a small number of images per class and the number of classes is fairly large. The CIFAR10 and CIFAR100 datasets contain 60,000 images each. Images in both datasets are 32x32x3 RGB color images.

In CIFAR-10, there are 10 classes and each class has 6,000 images. The dataset is divided into 50,000 training images and 10,000 test images. The following is the list of classes and some random images from each class from the CIFAR-10 dataset, so you can see what it looks like:

CIFAR-100 has 100 classes and 600 images per class. These 100 classes are divided into 20 superclasses. Each image has a fine label (the class that it belongs to) and a coarse label (the superclass it belongs to). The list of classes and Superclasses in CIFAR-100 is available at https://www.cs.toronto.edu/~kriz/cifar.html. Increasing the number of classes from coarse (20) to fine (100) can be helpful to maximize inter-class variability. This means that we want our model to consider two similar-looking objects in the image to belong to a different class. For example, a bed and a couch look similar but not exactly the same, and placing them in separate classes will ensure that they look different to the trained models.

The algorithm evaluation process for CIFAR is the same as in ImageNet. The best reported top-1 error for CIFAR-10 is 3.58%, and for CIFAR-100 is 17.31%, as reported by Saining Xie et al. in Aggregated Residual Transformations for Deep Neural Networks, where they presented the novel ResNeXt architecture. The Current state-of-the-art techniques for image classification using Deep learning results on CIFAR-10 and CIFAR-100 can be found at http://rodrigob.github.io/are_we_there_yet/build/classification_datasets_results.html and https://github.com/RedditSota/state-of-the-art-result-for-machine-learning-problems.