Create a lightweight thumb shell for one slide. The clone is * materialized later by the IntersectionObserver. Event handlers * look up th

Simple rule-based fallback when TCN weights are unavailable. Logic: - Wrist y-position above shoulders at swing peak -> Se

Return source duration in seconds via ffprobe.

Confirm the processed video is actually in `results` before nuking raw.

Re-clone a single thumb (live-update path). No-op if the thumb * hasn't been materialized yet — it'll pick up current content when * it sc

Resample a (T, D) array to (target_len, D) via linear interpolation. Matches the interpolation logic in extract_pose_sequence().

Infer camera_id from the video filename. Matches filenames containing 'Court' followed by a court number (e.g. StMarys_Court2.mp4). Retur

Try to extract player names from a plain HTML table.

Pre-label bounces into <clip>.gt.json. Returns count of seeds written.

Pick n clips stratified across source_video and velocity bucket.

Re-encode src (mp4v) → dst (H.264). Returns True on success.

Upload via curl (avoids miniforge arm64 SSL issues). Returns True on success.

Append new manifest rows, writing header if file is new.

Project API shots onto the top-half (opponent's side) court for ShotMap. * * Mirror by **bounce location**, not by who hit. Previously the code mirr

Derive spacing shots for the bottom-half spacing viz. * Only includes shots with both player + ball at contact. Quality classified by * distance:

Build the player-coverage occupancy grid for the bottom half-court. For every frame, project each detected player's foot position into court

Direction-only error breakdown attributed to the near (P1) player. Tennis convention: an error belongs to the player who failed to make the s

Load keypoints and compute feature vectors for all valid entries. Returns: X: (N, N_FEATURES) feature matrix y: (N,) integer

Net-approach detector + heuristic outcome. Approach: near player's foot svg_y crosses the service line (svg_y = 60) and stays in front of it

4-zone court-position breakdown for the near player (P1), per the `coach-insights-spec.md`: inside baseline / on baseline / 5-10 ft behind / 1

Build the per-shot record used by the frontend courtmaps. Bounce coordinates come from compute_bounce_positions() (the single source of truth

Group shot frames into rallies. A gap > gap_seconds between consecutive shots is treated as the start of a new rally.

Classify each detected bounce as in-bounds or out-of-bounds using the court-space coordinates obtained via homography projection.

Write a label for one clip to the DB.

Bulk-upsert all rows from a manifest CSV into swing_labels. With `preserve_labels=True` (default), any row that already has a non-empty label

Main detection pass + a second focused pass on the far-court region. The far player is typically only 30-60 px tall in a 1080p zoome

Detect frames where a player hits the ball by finding y-direction reversals in the ball trajectory. Args: ball_track: list of (x

Draw a fading ball trajectory on the main frame. Parameters ---------- frame : np.ndarray Current video frame (modified in-p

Draw tennis court lines as a semi-transparent white overlay. Keypoint dots and debug labels are intentionally omitted.

Draw ball trace and accumulated bounces on the minimap. minimap: reference court image (will be modified in-place) homography_inv: frame

Draw player positions on the minimap using foot position projected through homography. Each player gets a filled dot in their assigned color

Draw clean player bounding boxes with a small pill label. color is ignored when player_dict has multiple players — each gets its own col

Draw stroke classification label above each player's bbox when a swing is active at this frame. Args: frame: Video frame (numpy

Resize video to target resolution using ffmpeg (GPU NVENC when available).

Extract a flat feature vector from a (T, 17, 3) keypoint sequence. Feature groups (~45 total): - Wrist velocity stats (6): max/mean/st

Extract a fixed-length pose keypoint sequence for one player in a window. Args: handedness: 'right' (default) or 'left'. Lefty seque

Generate bounce dot map and player shot-position dot map on a top-down court. Bounce heatmap: color-coded dot per detected bounce (stroke ty

Call GPT-4o-mini to generate a grounded scouting report. Grounded in the real coach-insight calculators (position / net / error / rally summa

Determine the best homography matrix from detected court points. Parameters ---------- points : list[(x, y) | None]

Detect court keypoints, returning list of 14 (x, y) tuples.