This is a YOLO3 classifier for Japanese coins. The following dataset was used to train the classifier:
https://drive.google.com/open?id=1-_JkQ7E8N0toHYVHFKTuzugOg_OQMfEO
https://drive.google.com/open?id=1GUXLuv95ixrZYHcpM0JnWHUHoc_hk--k
https://drive.google.com/open?id=1n-4PwJUWzmUQuHnLwhnzQeVrYt_qrn_z
https://drive.google.com/open?id=1YjCo09Rh9pdtU44jR3eoG8GqmVchxhj_
https://drive.google.com/open?id=1SmHncEczhsGXXTBUUsaVFUPBVh-PWWF1
https://drive.google.com/open?id=1PMs6Q3Z08skZWCAmjBtd8-Sff9R_I6Z5
It contains 6 classes (1 - 5 - 10 - 50 - 100 - 500 Yen). Each class has 300 images. The data is split into 90% training and 10% for testing.
The dataset had been built from Google Images using the script from PyImageSearch
It had been filtered manually, then Augmentor library was used to increase the of images for each class (by zooming, cropping, rotating, etc..)
Finally, the data was labeled using BBox-Label-Tool
The training procedures was implemented on Google Compute Engine instance with Tesla K80 (to use CUDA for training).
The latest weight file (10000 iterations) can be found on this link:
https://drive.google.com/open?id=1nLIT9OwG00tB40QU52_mb-b1dLJ6DujM
It has average loss of 0.2, although it's still needs more training, as it can barely draw the bounding rectangles (in the previous weight file, it couldn't even identify or draw bounding boxes). Please feel free to fork the repository and continue the training.