set-solver

Solve the SET card game using OpenCV. Takes an image of a game of SET, returns same image with the sets indicated by drawing colored boxes around them.

Here's how it works.

Setup

1. Install OpenCV

Mac

Linux

2. Install python libraries:

pip install -r requirements.txt

Or use a virtualenv if you don't want to clutter your global packages

python -m virtualenv venv
source venv/bin/activate
pip install -r requirements.txt

3. Solve SET!

To solve a game and display the image with boxes around the sets:

./solve_set.py [filename] --display

Full usage:

usage: solve_set.py [-h] [--game GAME_NUM] [--write] [--display] [filename]

Solve SET from a game image.

positional arguments:
  filename         Game image filename

optional arguments:
  -h, --help       show this help message and exit
  --game GAME_NUM  use a test image from image-data/set-
                   games/setgame<GAME_NUM>.jpg
  --write          Write the solved image to solve-out/solved.jpg
  --display        Display the solved image with cv2.display()

Files

• image-data/ - All image data, including raw game images, labeled card images.
• vendor/ - where the Noteshrink code (for color bucketing) lives.
• avg_colors.py - Single use script to get the average shape color values from each of the red, green, purple images.
• card_finder.py - Given a game image, outputs images of all cards found.
• classes.py - Classes representing set games and cards.
• classify_card.py - Given a card image, outputs the best guess of what card it is.
• classify_card_accuracy.py - Rate how well classify_card.py does against a directory of labeled card files.
• common.py - Common constants or functions shared between scripts.
• process_card.py - Process a card image so that it's more easily classified by classify_card.py.
• extract_shapes.py - Cut out one to three shapes from a card image.
• label_all_cards.py - Single use script to easily generate labeled cards.
• solve_set.py - Script that runs the whole pipeline - takes in a game image file and displays that image with the sets overlaid.
• test.py - Tests for each chunk of the pipeline.

Future tasks

• Increase card classification accuracy - pretty good, but not perfect yet
  • "Shove a neural net into it" - optional if OpenCV isn't enough (probably not necessary, but could be fun)
    • I don't want to take hundreds of pictures of cards, so maybe fake a training set? Take the same image and artificially introduce jitter in a variety of ways (position, skew, rotation, white balance, lighting, etc) that mimics the real differences we'd get
• Better than brute force way to solve SET? Might be interesting to think about if SET's # cards on table, # attributes, # categories per attribute were increased
• More tests in general
• Make it run on a phone
  • React Native app that sends an image to a Flask app?
  • Have the whole thing run on the phone? Going to require an entire rewrite in Java or something

Why?

The full story is here, but because SET is fun, computer vision is awesome (and so is OpenCV), and I needed something to do at the Recurse Center.