Thank you for the advice (I'm new to this). The model performs better with a dark background, for example white dice on a wood table. I trained it with differ

Thank you for the advice (I'm new to this). The model performs better with a dark background, for example white dice on a wood table. I trained it with different colors of dice in the hopes that that would help it generalize. The model also performs better with top-down images of dice. I did train with a lot of oblique angles of dice (3000 images in total). That helped somewhat but it is still far from ideal. The model currently recognizes a standard D6 dice with pips. I made a separate model that recognizes D6 and the boat / castle die from Catan: Cities and Knights. That model is only used with the associated game type. Classifying dice rolls is an interesting problem to me because the same "orientation" of the die might translate into different results of what is the actual value. The model has no concept of what "up" is. The oblique angles issue also adds complexity. In my current approach, I opted to train the model to recognize an entire die. Another thought I had was to train it to recognize the individual faces as well as the tabletop. Then if I could find the face that was furthest away from the tabletop, I could predict that as the value. One last problem which I haven't solved yet is the concept of re-rolls. The detection is somewhat simplistic in that it looks for N consecutive rolls with the same result. It requires an empty frame where nothing is detected to "prime" the detector for the next roll. This was done because I didn't want the detector to redundantly record the same image of dice as multiple rolls. I drew a lot of inspiration for this project from the following tutorial: https://github.com/nell-byler/dice_detection https://github.com/nell-byler/dice_detection. Nell also has a Dice Dataset on Kaggle: https://www.kaggle.com/nellbyler/d6-dice https://www.kaggle.com/nellbyler/d6-dice