A computer vision project for playing the game of Dobble automatically
The project is explained in the dissertation write-up provided in the repo.
The basic idea of the project is to automatically detect cards within the video frame, crop them, then identify the 8 different symbols on each card.
The orignial card images were aquired from Leo Reynolds on Flickr
The cards were then manually annotated using labelImg
The datasets are then synthesised using openCV with a collection of background images.
The bounding box and multiclass classifier models are created using fastai, a library built on top of pytorch.
A tkinter app loads the models and visualises predictions with a GUI.
The system can run at over 20 frames per second, correctly identifying the matching symbols between cards around 90% of the time on real world examples.