For chessboard grid detection:

• X-corner detection using Automatic Chessboard Corner Detection Method
• Iterative grid fitting from this

For piece detection:

• Data generation using LabalImg (a lot of time spent)
• Fine tune pretrained faster_rcnn_resnet50_coco from object detection api

Files:

• detection.py : piece detection
• main.py : X-corner detection, grid-fitting, integration with piece detection
• gui.py : runnable file to use this pipeline
• test.py : bulk testing of images in test folder
• ...

• Python 3.6
• OpenCV3
• tensorflow including tensorflow's models repo
• protoc
• PIL
• numpy
• scipy
• svglib
• reportlab
• pandas
• ...

git clone [email protected]:jonkoi/chess_detection.git
cd chess_detection
git clone [email protected]:tensorflow/models.git
cd models/research
protoc object_detection/protos/*.proto --python_out=.
set PYTHONPATH=PATH\TO\models\research;PATH\TO\models\research\slim (for windows)
cd ../..

• Download my fine tuned model and optionally test images (not involved in training). Extract them in the main repo folder.
• Train dataset

For simgle images:

python gui.py

For bulk testing of test images:

python test.py

Here is video of the test.

The piece detection module uses deep learning (not mobilenet) which can make inference very slow. The process is sluggish even with a GTX860M.