cogsys-tuebingen/blurball ? reverse-engineered prompt

Build me a Python project that can detect a table tennis ball in video and also estimate the motion blur around it, including the ball position, blur direction, and blur length.

I want it to work with a trained model on a dataset of table tennis clips, and I want clear commands for training, evaluating, and running inference on a video. If a video has repeated frames, handle that automatically so the results stay stable. Please make it easy to point the app at my dataset folder and at a model weights file, then run the detector from the command line.

Use the existing project structure if it helps, and keep the setup simple with one requirements install step. Add or update the README so I can follow it without guessing. If you need to check current library docs or best practices, go ahead and look them up online.