Context I created this dataset to teach the basics of fully convolution networks for semantic segmentation of images. Most real-world semantic image segmentation tasks require building huge networks that are slow to train and experiment with. The dataset was generated by selecting up to 3 random 28px x 28px grayscale images from the [MNIST dataset](https://www.kaggle.com/c/digit-recognizer/data) and copying them in to a single 64px(height) x 84px(width) image. The digits were pasted so that they did not overlap and no transformations were applied to the original images, so digits in M2NIST maintain the same orientation as the have in MNIST. Content The dataset has 5000 multi-digit images in `combined.npy` and 11 segmentation masks for every image in `segmented.npy`. The files can be read in using `numpy.load()`, for example, as `combined=np.load('combined.npy')` and `segmented = np.load('segmented.npy')`. The data in `combined.npy` has shape `(5000, 64, 84)` while the data in `segmented.npy` has shape `(5000, 64, 84, 11)`. Every element in `combined.npy` is a grayscale image with up to 3 digits. The corresponding element in `segmented.npy` is a tensor with 64 rows, 84 columns and 11 layers or channels. Each layer or channel is a binary mask. The k-th layer (0<=k<9) has 1s wherever the digit k is present in the combined image and 0s everywhere else. The last layer k=10 represents background and has 1s wherever there is no digit in the combined image and 0's wherever at pixels where some digit is present in the original image. Acknowledgements This dataset is ultimately derived from the data published by YanLeCun and his group http://yann.lecun.com/exdb/mnist/index.html and is licensed under the [Creative Commons Attribution-Share Alike 4.0 license](https://creativecommons.org/licenses/by-sa/4.0/). The code used to generate this dataset is available on [github](https://github.com/farhanhubble/udacity-connect/blob/master/segmented-generator.ipynb) as an Ipython Notebook. Inspiration The M2NIST dataset is released in the hope that it enables users to understand semantic segmentation and perhaps give us more insights about what neural networks learn and in turn leads us to smaller and more robust networks.