The Quick, Draw! dataset

Four formats, four purposes

The dataset is published in several variants, and the choice determines how much preparation is left to you:

• Raw — the original recording with absolute coordinates and per-point timestamps. Drawing speed exists only here.
• Simplified — normalised: aligned to the top-left corner, scaled to 256 × 256, resampled at uniform spacing and stripped of redundant points using the Ramer-Douglas-Peucker algorithm with epsilon 2.0. Timing information is dropped.
• Binary — the same simplified data in a compact form, considerably faster to read than line-delimited JSON.
• NumPy bitmaps — ready-rasterised 28 × 28 greyscale images for anyone who wants to start immediately.

DrawLa uses the simplified stroke format and does its own rasterisation at 96 × 96 pixels. The ready-made 28 × 28 bitmaps would have been more convenient but are too coarse: at that resolution bicycle spokes, window frames and animal legs merge into undifferentiated blobs. Rasterising yourself also keeps control over one thing that matters — training and inference using exactly the same procedure. See QuickDraw with PyTorch.

The most important quirk: abandoned drawings

This one is easy to overlook, and it shapes the entire dataset. In the original game, a round ends the moment Google's model guesses the word. Draw a very recognisable cat and you are cut off after three strokes. Struggle, and you draw for the full twenty seconds.

The dataset therefore systematically contains barely recognisable drawings — and hardly any carefully finished ones. A model trained on it inherits that expectation. This explains something DrawLa players notice regularly: put in effort and add detail, and you are often recognised worse than someone throwing down three rough strokes. The model is not broken; it learned what its data contained.

Practical consequence for play: the silhouette plus one or two characteristic features is enough. Anything beyond that moves your drawing away from what the model knows.

What the metadata holds

Every entry carries more than strokes and a label. Two fields are particularly useful.

recognized records whether Google's own model identified the drawing during the game. You can filter the set with it — at the cost of then training on another model's judgement and inheriting its blind spots.

countrycode gives the origin. It makes cultural differences visible that are by now well documented: circles are drawn in different directions in different countries, and houses and chairs follow regionally different conventions. For a model that reads the sequence this is a genuine source of error — for an image-based CNN, stroke direction is irrelevant.

Timestamps and a per-drawing identifier round it out. The dataset is published under CC BY 4.0, so it is free to use with attribution, including commercially.

Data quality: what to expect

These are millions of unvetted drawings by volunteers, with everything that implies. Some are simply mislabelled because people drew something else. Others are scribbles, letters or jokes. Some entries break off after a single stroke.

This is why label smoothing is not a finishing touch on this dataset but an appropriate measure: it stops the model committing to noisy labels with full confidence. Equally important is a realistic expectation of achievable accuracy — part of the remaining error is not fixable, because the source material itself is ambiguous.

On top of that comes genuine ambiguity between categories. Some class pairs are practically indistinguishable as quick scribbles: snail and seashell, cloud and sheep, blueberry and a plain circle. No model will separate those reliably, because people cannot either.

How much data do you actually need?

Using all 50 million drawings sounds tempting but rarely pays. DrawLa uses 5,000 drawings per category, around 1.7 million images in total. Beyond that order of magnitude the gain for a model with two million parameters is small, while download time, storage and training duration grow linearly.

The second reason is balance. The full dataset is markedly imbalanced — popular categories hold many times the drawings of rare words. A fixed cap per class restores balance without needing class weights in the loss function.

What it means for DrawLa

DrawLa adopts the 345 categories as its vocabulary and classifies rasterised sketches with a compact CNN. The dataset therefore supplies both: the word list drawn at random during play, and the basis on which recognition works at all.

Why stroke sequences become images rather than staying sequences is covered in RNN versus CNN. The implementation is described in QuickDraw with PyTorch, and Google's own component for fetching individual doodles is discussed under the GitHub implementation.

DrawLa.app is independent of Google. Quick, Draw! and the dataset are works of Google Creative Lab, named here as a data source; Google does not endorse or sponsor DrawLa.