Central configuration for the tennis analysis pipeline

Detects ball bounce frames using a trained CatBoost regressor.

Canonical tennis court reference in pixel space. Used for homography mapping and minimap visualization.

Temporal Convolutional Network for tennis stroke classification. Input shape: (batch, seq_len, 34) Output shape: (batch, 4) -- raw logi

Computes the homography between the canonical tennis court reference and the image frame using detected court keypoints.

Loads keypoint sequences from .npy files and applies the augmentation pipeline. Returns (seq_len, 34) tensors after hip-center normalization.

Detects tennis swing events from per-frame pose keypoints. Usage: detector = SwingDetector(velocity_threshold=15.0, ball_proximity=3

Stroke recognition model

1D U-Net trajectory inpainting model from TrackNetV3.

Time sequence model for stroke classifying

High-level wrapper around StrokeTCN for use in the pipeline. Falls back to a rule-based heuristic when weights are unavailable.

Fills in missed ball detections (None entries) using InpaintNet. Usage: rectifier = TrajectoryRectifier(weights_path, frame_w, frame

One TCN residual block with dilated causal convolutions.

* Recordings list. Ported from docs/brand-drop/mocks/matches-list.html. * * Layout: * - h1 "Every recording, scrubbable." (clay italic on "scrubb

Phase 2: verify choose_and_filter_players returns 2 players per frame.

Phase 1: verify raw YOLO detections see both players.