ProClone: Your NBA Shooting Posture Mirror

ProClone is a lightweight, phone-video–based system that tells you “You shoot like …” an NBA superstar. By extracting five key joint angles frame-by-frame, it matches your biomechanical signature against 20 legends—no sensors required.

---

🚀 Key Features

• Video-only input — just record on your phone
• Pose detection via OpenCV + MediaPipe (shoulder, elbow, wrist, hip, knee)
• Residual MLP classification
  • 90 % frame-level accuracy
  • 70 % shot-level majority-vote accuracy
• Nearest-neighbor search on z-score-normalized angle vectors
• Real-time inference: under 3 s per clip on a laptop GPU
• Top-3 matches so you know which pros you resemble

---

🎯 Why ProClone?

Most shooting apps point out errors; ProClone tells you who you shoot like—making practice more engaging, motivating, and personalized.

---

🛠️ How It Works

1. Capture & Preprocess

1. Record your jump shot on any phone camera.
2. Extract keypoint coordinates & compute five joint angles per frame.
3. Resample each clip to a fixed 34-frame sequence for consistency.

2. Pose Detection

We use MediaPipe’s Pose solution to detect 33 landmarks, then compute five joint angles:

import cv2
import mediapipe as mp
import numpy as np

mp_pose = mp.solutions.pose.Pose(
    min_detection_confidence=0.5,
    min_tracking_confidence=0.5
)

def extract_angles(frame: np.ndarray) -> dict:
    # Convert BGR → RGB
    results = mp_pose.process(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
    if not results.pose_landmarks:
        return {}

    lm = results.pose_landmarks.landmark
    def angle(a, b, c):
        a, b, c = np.array(a), np.array(b), np.array(c)
        ba, bc = a - b, c - b
        cosang = np.dot(ba, bc) / (np.linalg.norm(ba) * np.linalg.norm(bc))
        return np.degrees(np.arccos(np.clip(cosang, -1, 1)))

    pts = {p.name: (lm[p].x, lm[p].y, lm[p].z)
           for p in mp.solutions.pose.PoseLandmark}

    return {
        "shoulder": angle(pts["RIGHT_HIP"],    pts["RIGHT_SHOULDER"], pts["RIGHT_ELBOW"]),
        "elbow":    angle(pts["RIGHT_SHOULDER"], pts["RIGHT_ELBOW"],    pts["RIGHT_WRIST"]),
        "wrist":    angle(pts["RIGHT_ELBOW"],    pts["RIGHT_WRIST"],    pts["RIGHT_INDEX"]),
        "hip":      angle(pts["RIGHT_SHOULDER"], pts["RIGHT_HIP"],      pts["RIGHT_KNEE"]),
        "knee":     angle(pts["RIGHT_HIP"],      pts["RIGHT_KNEE"],     pts["RIGHT_ANKLE"]),
    }

3. Feature Normalization & Matching

Flatten per-frame angles → 170-D vector

Z-score normalize against NBA dataset

Cosine similarity for nearest-neighbor lookup

4. Inference & Feedback

Residual MLP predicts your top match (90 % frame-level)

Shot-level majority vote yields 70 % accuracy across full clips

Nearest-neighbor returns Top-3 “You shoot like …” recommendations

📦 Installation & Usage

git clone https://github.com/your-org/proclone.git
cd proclone
pip install -r requirements.txt

from proclone import ProClone

pc = ProClone(
    mongo_uri="mongodb+srv://<user>:<pass>@cluster0…",
    db_name="pose_db_new",
    n_frames=34
)

Analyze your shot (pass a DataFrame of frames):

matches = pc.match_user_shot("Jon", video_frames_df)
print(matches)
 [("Jalen_Green", 95.2), ("Chris_Paul", 94.7), ("Donovan_Mitchell", 93.8)]