asigatchov/ravel-vb-beachvolleyball-tracking

TAPe-VB, also referred to as RAVEL-VB in the checkpoint filenames, is a compact tool for analyzing volleyball video footage. It watches a video clip and detects where the players and the ball are in each frame, keeps track of individual players as they move across frames, and can produce a new video with those detections drawn on top so you can see the tracking visually. The project includes two ways to run the model. One path uses PyTorch, the machine learning library many AI projects are built on, and the other uses OpenVINO, a toolkit designed to run models efficiently on regular computer processors without needing a graphics card. Both come with example pretrained model files included in the repository, so there is no separate training step required to try it out. To use it, you install Python 3.10 or newer along with the listed dependencies, then run one of two provided scripts, either infer_torch.py or infer_openvino.py, pointing them at a video file. The tool writes out a JSON file containing the pixel positions of detected players and the ball, their confidence scores, and short term tracking IDs for each player, and it can optionally render an annotated output video as well. Several command line options let you adjust things like detection sensitivity and which computing device to use. The README notes some practical limits: the model can struggle with a very small or blurry ball, players blocking each other, sudden camera cuts, or footage that looks very different from what it was trained on. The bundled models are specifically built for beach volleyball style footage. No license is stated in the README, so anyone wanting to reuse or redistribute this code should check with the repository owner directly.