// the find
vietnh1009/QuickDraw
Implementation of Quickdraw - an online game developed by Google
A PyTorch CNN that replicates Google's Quick Draw game — you draw in front of a webcam using a colored pen, and the model guesses what you drew. Covers 20 of the 345 Quick Draw dataset categories. Good for learning how to hook a trained classifier up to a real-time OpenCV pipeline, not much beyond that.
The color-tracking pen detection with OpenCV is a clever low-fi alternative to a drawing tablet — works without any special hardware. Training code is clean and minimal, easy to follow for someone learning PyTorch image classification. The repo ships a pretrained model so you can run it immediately without downloading and preprocessing the npz dataset files yourself.
Only 20 of 345 available classes with no explanation of why those 20; extending it requires manually figuring out the dataset download and config change yourself. Last touched January 2023 and pinned to Python 3.6, which is well past EOL — expect dependency resolution pain with modern PyTorch. The 'drawing app' mentioned in the README is listed as 'released soon' and never shipped. Color-blob tracking for pen detection is brittle in anything other than controlled lighting.