Issue #12: App development and pipeline integration

Endpoint alternative chosen: Gradio tab inside the existing app.py.

Input: one video file. Output: annotated cut 2D video, 3D skeleton animation video, keypoints CSV, and good/bad classification JSON.

A14: Advanced Exercise Pipeline

Features: Automated 'Ugly' recording rejection + 'Good/Bad' form classification.

A15: Exercise Scoring (0–4 regression)

Score scale: 0 = perfect form, 4 = worst kept clip.

Bands:

• GREEN < 1 — acceptable form
• AMBER 1–2 — borderline, consider another take
• RED ≥ 2 — poor form

The same upstream pipeline as A14 is reused (pose extraction + 3D lift + A12 start/stop cut). Decision-time of the NN and the overall response-time breakdown are reported alongside the score.

A16 — Final unified endpoint (3D alternative)

Record a clip with your webcam (or upload one), then click Run A16 endpoint. The result appears on the right: a video with the skeleton overlaid on your recording, plus the full Part-II chain output — pose → PoseNet→Kinect 2D → 2D→3D → start/stop cut → ugly/good-bad → 0–4 score.

Processing is currently CPU-only and runs the full chain end-to-end, so a 5-10 s clip can take roughly 20-60 s on the HF Space.

MediaPipe 3D Pose Livestream

Live webcam pose estimation using MediaPipe Tasks (pose_landmarker_lite.task). The left panel shows the 2D skeleton overlay; the right panel shows the 3D world landmarks.