Fashion MNIST

Fashion-MNIST is a modern drop-in replacement for MNIST. Released by Zalando Research in 2017, it packs 70 000 tiny grayscale images of apparel—sneakers, shirts, coats—into a lightweight benchmark. Its familiar format keeps setup trivial, while richer visuals pose a tougher challenge.

Zalando’s quality-control cameras captured millions of 96 × 96 product shots. Han Xiao et al. down-sampled these to 28 × 28, grouped them into ten balanced classes, and open-sourced the result. The idea: upgrade MNIST difficulty without touching loaders or evaluation scripts.

Digits look alike; CNNs now exceed 99.7 % accuracy, offering little headroom. Clothes introduce texture, deformation, and subtle inter-class overlap, dropping baselines below 90 %. By swapping only pixel data and preserving the gzipped uint8 files, Fashion-MNIST raises the ceiling while keeping APIs identical.

• Train / Test: 60 000 / 10 000 images
• Classes: T-shirt/top, trouser, pullover, dress, coat, sandal, shirt, sneaker, bag, ankle-boot
• Resolution: 28 × 28 px, 1 channel, 0–255 grayscale

Fashion-MNIST powers tutorials, quick papers, and embedded demos where CIFAR-10 or ImageNet would be overkill. It underpins research on adversarial robustness, quantization, federated learning, and tiny-ML hardware. Tougher sets exist—STL-10, CIFAR-100, TinyImageNet—but Fashion-MNIST remains the rapid sanity check for new vision pipelines: the “next-gen drosophila” of machine learning.