🏓 Ping Pong Sport Analytics

Computer vision pipeline for table tennis analysis — player tracking with persistent identity, ball tracking, and bounce detection.

⚠️ This repository showcases the methodology and results. Source code and trained models are not included.

---

Pipeline

Video Input
    │
    ├── Players ──── YOLO (fine-tuned) → ByteTrack → OSNet ReID
    ├── Ball ─────── YOLO (fine-tuned) + CVAT correction for demo
    ├── Table ────── YOLO-Seg (fine-tuned)
    └── Bounces ──── scipy.signal.find_peaks + table polygon filter

How It Works

Player Re-Identification — ByteTrack handles frame-to-frame tracking. When a player exits the frame and re-enters (or switches sides), OSNet compares appearance embeddings against a gallery of known players to restore the correct identity. A position-based fallback and anti-duplicate enforcement guarantee stable IDs throughout the video.

Ball Detection — A fine-tuned YOLO model detects the ball. Missing detections were pre-annotated by the model then manually corrected in CVAT to achieve full coverage for the demo.

Bounce Detection — Peak detection (find_peaks) on the ball's Y-coordinate trajectory, filtered by whether the bounce point falls within the table polygon (cv2.pointPolygonTest).

Table Detection — A fine-tuned YOLOv8-seg model detects the table surface. For the final output, a 4-corner polygon calibration is used.

Stack

Component	Tool
Player detection	YOLOv8n (fine-tuned on 2.3k images)
Player tracking	ByteTrack
Player ReID	OSNet x1_0 (torchreid)
Ball detection	YOLOv8n (fine-tuned on 15k images)
Table detection	YOLOv8n-seg (fine-tuned on 1K images)
Bounce detection	scipy.signal.find_peaks
Annotation	CVAT

Future Work

• Homography-based mini-map for top-down tactical visualization
• Serve detection and point scoring
• Player movement heatmaps

---

Built by @IIIllllIlIlllII