It probably a flaw in the model. Thank you for the videos, I will add some of those frames to my training data, it most likely fix the issues, it may take a little time to train it tho.
