googlecreativelab/quickdraw-dataset

The Quick, Draw! Dataset is a collection of 50 million hand-drawn sketches contributed by players of Google's Quick, Draw! game, where participants had 20 seconds to draw a prompt and a neural network tried to guess what they drew. The drawings span 345 categories ranging from everyday objects to abstract concepts. Each drawing is stored as a sequence of pen strokes with x/y coordinates and timestamps, along with metadata including the category, the player's country, and whether the game successfully recognized the drawing. Google has released the data in several formats to suit different use cases. The raw format stores each drawing as a JSON record in a plain text file, one drawing per line. A preprocessed version cleans up the strokes, removes timing data, and scales everything into a standard 256x256 pixel region. There is also a binary format for faster loading, and a set of 28x28 pixel grayscale bitmap images in NumPy format for anyone who wants to treat the drawings as images rather than vector paths. All formats are hosted on Google Cloud Storage and can be downloaded by category. The raw data for each category arrives as a separate file, and the download can be done with a single command using Google's gsutil tool. A subset of the data, 75,000 samples per category, was prepared specifically for training the Sketch-RNN model, a generative model that can produce new drawings in the style of a given category. That version is stored in compressed NumPy files and was used in research on teaching computers to draw. The dataset is made available for developers, researchers, and artists. Google notes that while the drawings were individually moderated, the collection may still contain some inappropriate content.