A local QuickDraw clone that judges your pen through a lens

What it does

This is a self-contained PyTorch re-creation of Google’s QuickDraw game. You hold a blue, red, or green pen in front of your webcam; OpenCV tracks the tip with a yellow circle, and a CNN trained on 20 doodle categories guesses your sketch. A canvas-based drawing mode is promised in the README but not yet released.

The interesting bit

The input pipeline relies on color detection rather than a touchscreen or mouse, turning any colored marker into a stylus. It is a pragmatic, low-friction hack that avoids extra hardware.

Key highlights

• Trained on the first 10,000 samples from each of 20 Quick Draw dataset classes.
• Ships with a pretrained model stored at trained_models/whole_model_quickdraw.
• Runs fully offline after the initial model and data files are present.
• The category list is governed by a CLASSES constant in src/config.py, so swapping or expanding classes only requires the corresponding .npz files.

Caveats

• The canvas drawing app is still unreleased; only the webcam color-tracking mode is available.
• The README is sparse on model architecture, inference speed, or accuracy numbers beyond the loss curves.

Verdict

Worth a look if you need an offline demo of sketch recognition without cloud APIs. Pass if you want a polished UI, a broad vocabulary beyond 20 objects, or detailed documentation.