Natural image datasets

Image classification usually includes a wider range of objects and scenes than the MNIST handwritten digits. Most of them are natural images, meaning images that a human being would observe in the real world, such as landscapes, indoor scenes, roads, mountains, beaches, people, animals, and automobiles, as opposed to synthetic images or images generated by a computer.

To evaluate the performance of image classification networks for natural images, three main datasets are usually used by researchers to compare performance:

• Cifar-10, a dataset of 60,000 small images (32x32) regrouped into 10 classes only, which you can easily download:

  wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz -P /sharedfiles
  tar xvzf /sharedfiles/cifar-10-python.tar.gz -C /sharedfiles/

  Here are some example images for each class:

  

  Cifar 10 dataset classes with samples https://www.cs.toronto.edu/~kriz/cifar.html

• Cifar-100, a dataset of 60,000 images, partitioned into 100 classes and 20 super...