Wildlife MNIST

The Wildlife MNIST dataset contains MNIST digits with colored backgrounds and foregrounds with annotations, suitable for benchmarking disentangling or factor identification. Originally used for the project https://github.com/vitskvara/sgad. There are two versions - non-mixed and mixed. In the non-mixed version (data.npy and label.npy), the background and foreground textures are the same for all digits of a single MNIST class, therefore only a single label describes each sample. In the mixed version (data_test.npy and labels_test.npy), each sample image has a random digit, background and foreground (out of 10 classes for each factor of variation). Then, the label is a tuple of three numbers, describing the individual (digit,background,foreground) labels. Note that the data is scaled to the interval [-1,1], so rescaling them by computing "x*0.5 + 0.5" is necessary for some applications that require them to be in the interval [0,1]. Example images from both versions of the dataset are included. Note that the dataset was originally used in "Sauer, Axel, and Andreas Geiger. Counterfactual generative networks. 2021."