Model Choice

After researching various OBB models, we chose to use the YOLOV5 OBB model. It appeared to be the most mature option, with comprehensive documentation and impressive performance. You can find the YOLOV5 OBB repository here.

Dataset Issues

The initial dataset we used was too small, containing fewer than 100 photos and cards. With this dataset, we were unable to achieve satisfactory results. To quickly gather more data, we created an artificial dataset that combined images of cards and weapons. We used cropped images of cards and weapon images from the classification dataset.

Training

After various training sessions and experiments, we found that the following requirements were necessary for achieving good results:

• Use a pretrained model for quicker convergence.
• A small batch size (1 or 2) yields better results.
• Prefer the Adam optimizer.
• Do not use flip data augmentation, as it interferes with the training.

Results

Here are the results we obtained:

The cards are detected with precision.