koisland/goat

Go Automatic Tallier

Go Automatic Tallier (GoAT)

Automatically tally score for Go boards from an image.

---

Table of Contents

• Getting Started
  • Conda
  • Docker
• Usage
• Scoring
• Imaging

---

Getting Started

git clone -b main [email protected]:koisland/GoAT.git
cd GoAT

Conda

conda create --name GoAT python=3.10
conda activate GoAT
pip install -r requirements.txt

Docker

docker build . -t go_at:latest

---

Usage

Currently, usage is limited to the command-line with main.py as the entrypoint.

By default, black is marked as 1.0 and white is 0.0.

usage: main.py [-h] -i INPUT -s SCORING [-k] [-cb CAP_BLK] [-cw CAP_WHT]

Calculate score from a Go board image.

options:
  -h, --help            show this help message and exit
  -i INPUT, --input INPUT
                        Input image.
  -s SCORING, --scoring SCORING
                        Scoring method.
  -k, --komi            Apply komi.
  -cb CAP_BLK, --cap_blk CAP_BLK
                        Captured black stones by white.
  -cw CAP_WHT, --cap_wht CAP_WHT
                        Captured white stones by black.

For example, this command reads docs/images/9_9.png, a digital image of a board and scores it using Chinese scoring with komi applied to White.

Conda

python main.py -i docs/images/9_9.png -s Chinese -k

Docker

docker run go_at:latest -i docs/images/9_9.png -s Chinese -k

{'Black': 44, 'White': 44.5}

---

Scoring

Both Chinese and Japanese scoring methods are available with komi as a toggle-able option.

By default, komi is set to 7.5 for Chinese scoring and 6.5 for Japanese scoring.

Japanese scoring is a WIP.

• Currently can handle scenarios without seki.

Imaging

Accomplished through use of packages:

• opencv-python
• numpy
• statistics.

General workflow is as follows:

1. Gaussian Blur ->
2. Global Threshold (Two for each piece type)->
3. Distance Transform (Only black) ->
4. Contour Detection ->
5. Cluster Pixels ->
6. Map Pixels to Dimensions ->
7. Generate Board

Several assumptions are made about the --input image.

• The image that only encompasses the digital board
  • A real-world image would require manually trimming such that the image only contains the board.
  • Lighting conditions are another issue that could be handled with localized histogram equalization with cv2's clahe.