Expansion of Dataset: This is perhaps the single most important improvement
that could be made to our current model. Most of our issues (not working well with multiple
cards in the same image, struggling with non-standard cards, etc.) could be improved or perhaps
even resolved by simply creating an even larger and even more robust dataset. In addition, the other
28 cards in a standard deck could also be added to the model in order to change its use over to
identifying poker hands.
Improvement of Generality: Expanding the image library to include more non-standard
card images would likely allow the model to generalize better across card types it has never encountered
before. This may be a monumental task, however, with how many different variations of cards there are.
Recommendation System: A useful additional feature of the web application would be allowing
the ability for the user to receive suggestions based on their current hand and even the current state of the
game (as input by the user).
Experiment with Different Models: We primarily focused on maximizing the precision of a
single model throughout our training process (specifically the faster_rcnn_resnet101 model). Exploring
different models may offer advantages over this model.
Closed Loop Features: An extremely cool addition to the web application would be adding
the ability for users to provide feedback about not only if an identification is incorrect but also if an
identification is missed. The user could then draw a box around any mislabeled objects and the model would
use this input to further train itself.