CourtVision AI
- When
- Dec 2025 to Present
- Result
- Test accuracy 73%
- Python
- PyTorch
- Hugging Face
- NBA API
- Pandas
Context
Basketball has been my sport since I was 14, from my local club to the Irish national squad. CourtVision AI is where that meets machine learning: can a model watch a clip and tell you what kind of shot it just saw?
What I built
- NBA game footage as training data, with the NBA API and Pandas in the data pipeline.
- A fine-tuned version of Facebook's X3D-M, a pre-trained video recognition model, built with PyTorch and Hugging Face.
- A classifier that sorts each shot into one of three classes: free throw (1 pt), field goal (2 pts) or three-pointer.
Results
The fine-tuned model reaches 73% accuracy on the held-out test set.
What's next
The project is still in progress. The code isn't public yet; a repository will follow.