Spoodle, short for speed doodle is a game that I wanted to create based on my love for fun but simple party games such as skribble.io. The goal of Spoodle is in a short period of time, create the best possible drawing of a given prompt, and accuracy is judged by a neural network. How close the neural network believes the drawing is to the prompt, the higher the score, and these scores are used to compete against other players and friends.

There will be four phases to this project.

1. Use a pretrained model to create a program that can classify doodles.
2. Write the software that allows players to create doodles.
3. Integrate the classification and doodle programs into a website that users can interface with.
4. Possibly train a neural network with a dataset to provide another classification option.
5. Release, get recommendations on improvements, and continue to improve Spoodle.

The first phase, using a pretrained model to classify doodles, will be relatively simple. I will be using DoodleNet, a pretrained convolutional neural network (CNN) that is trained on Google's Quick Draw dataset. There are two CNN classifiers in the repo, one running with tf.js and one running with keras. Since I am already familiar with Keras in python, I am going to try to learn Javascript and use the tf.js model.

Needed to up the brush size:canvas size ratio to 1:28, since it needs to match as closely as possible to the dataset.

• Perhaps add a multiplier for an early submit, where the time left when submitted gives a higher score.
• Quick submit hotkey.
• Add percentage while you draw?
• After submitting, add points based on placing and show other peoples' drawings and their percentages.

https://www.youtube.com/watch?v=ABN_DWnM5GQ https://github.com/yining1023/doodleNet