Machine Learning for League of Legend mini-map

Technical Details:

General idea: Use node script to detect whenever the Mysterious Web Socket is open, listen to the incoming data and save the result as JSON file.

In order to run a supervised learning, another problem is how to label dataset:

solve the problem of organizing CSV files and JPG files:

Save the data into a .npz file which save all stuff as raw numpy arrays
During labelling, another problem is the data bias:

A misrecognition could happen if there is not enough data for a certain champ.

To avoid overfitting and underfitting for different champs: code to balance the dataset under the function check_champs

Focal loss could be used to balance the dataset
However, how to get the time which is labelled timestamps:

OCR: use computer vision to get the timestamp (YOLO), another option is Google Cloud Vision API (but need paying)
data augmentation could be a good method (flip the frame)but may cause confusion of the model (also flip the champs’ icon).
One limitation is that the RAM is supposed to storage many images at once.