The Quick, Draw! dataset: how a game became AI training data

For an AI to recognise doodles, it needs countless examples. One of the largest open sketch datasets came about in a charming way: through a browser game that millions of people played voluntarily.

A game collects data

In 2016 Google released the game “Quick, Draw!” as part of its “A.I. Experiments”. The task: you get a word and have 20 seconds to draw it while an AI guesses. What began as an entertaining experiment became, as a side effect, a gigantic data collection – because every drawing was stored (anonymised).

This gathered over 50 million drawings across 345 categories – from “apple” to “lighthouse” to “zigzag”. Google then made this dataset publicly available so that researchers and developers could work with it.

What's inside the data

The format is fascinating: what was stored is not just a finished image but the drawing as a sequence of strokes – that is, which lines were created in which order and direction. This means a sketch can be evaluated both as an image (each stroke drawn onto a surface) and as a temporal sequence (point by point).

Exactly this dual nature makes the dataset so valuable for research: it suits classic image recognition as well as models that handle sequences.

What it's used for

The Quick, Draw! dataset has become a standard playground for machine learning. Among the well-known results is Sketch-RNN, a model that not only recognises sketches but can draw new ones itself. Beyond that, it serves as a practice dataset in courses, for benchmarking models, and for creative projects around machine perception.

DrawClash stands in this tradition too: the recognition model distinguishes the same 345 categories. You'll find the complete list – in nine languages – under All words.

Why “many people” beat “perfect” drawings

The real treasure of the dataset is its diversity: because countless different people drew, it contains scrawly, crooked, minimalist and detail-loving versions of the same word. A model trained on it learns the shared essence of a subject rather than a single “correct” depiction – and thereby becomes robust against every player's quirks.

Sources & further reading

“Quick, Draw!” and its dataset are works by Google Creative Lab. DrawClash is an independent project and is not affiliated with Google.

Contribute a sketch yourself? At least give it a try!

Draw in DrawClash and see how the AI interprets your strokes.

Let’s Play