How joao-m-silva/padel_analytics works
Project type and purpose
• This is an open-source computer-vision analytics tool for padel match recordings. It ingests a raw video, allows manual court keypoint selection (homography), runs several trained vision models (player detection, pose estimation, classifier, ball tracker), computes statistics, and renders visualizations (heatmaps, velocity plots, 2D projections).
File structure
• Top-level orchestration:
• main.py / app.py – entry points.
• config.py – location of model weights, batch sizes, I/O paths.
• Core modules:
• trackers/ – individual tracker implementations (ball, players, keypoints) and a runner to chain them.
• analytics/ – code for physical projections, velocity/position statistics, aggregated data analytics.
• Helpers:
• utils/ – video I/O, unit conversions, homography helpers.
• constants/ – court dimensions, player-height data.
• UI:
• ui.py – interactive keypoint selector for court calibration.
• Examples and docs:
• examples/videos/ – sample inputs & demos.
• README, CONTRIBUTING, .github/ – contribution & issue templates.
README for additional insights
• States a fixed-camera assumption requiring 12 court keypoints, uses PyTorch models (pose/net/classifier), and produces a sequence: keypoint selection → tracking → analytics → visualization.
• Describes VRAM requirements, inference command (python main.py), and weight-download process.