50 million hand-drawn doodles across 345 categories, capturing the stroke-by-stroke creation process of millions of users. The dataset includes metadata such as country of origin, timestamps, and a boolean flag indicating if the drawing was recognized by the original classifier.
Use Cases
- Train a recurrent neural network (RNN) for sketch generation using the stroke sequences in the drawing column
- Analyze cultural differences in drawing styles by grouping visual features by the countrycode field
- Build a real-time sketch recognition engine using the 28x28 grayscale bitmap data
Strengths
- Contains 50 million drawings across 345 distinct categories including 'cat', 'bicycle', and 'airplane'
- Provides stroke-level vector data in .ndjson format and 28x28 grayscale bitmaps in .npy format
- Includes metadata fields for 'countrycode', 'timestamp', and a 'recognized' boolean flag
- Features simplified vector data that normalizes coordinates and removes pauses between strokes