cricket-rule-breaking

🏏 Cricket Rule-Breaking Detector — Front-Foot No-Ball (Google Colab)

One-line story: Fatima Noor loves cricket and built this Colab-only MVP to highlight front-foot no-balls on short videos with clear on-frame overlays.

---

Overview

This repository contains a Google Colab workflow to:

• Detect the bowler’s front foot near the popping crease.
• Produce an annotated MP4 with boxes/labels.
• Optionally add crease calibration and a LEGAL / NO-BALL verdict.

---

What I Used (Project Inputs)

• Model weights: best.pt (YOLOv8n fine-tuned for front-foot context).
• Videos: Short delivery clips sourced from YouTube (educational/research use). Clips were trimmed to ~6–12s where the front foot & crease are clearly visible.
• Training data: A small YOLO-format dataset focused on front foot near popping crease (curated and managed via Roboflow).

---

How I Trained the Model (Roboflow)

1. Create a Roboflow project and set the task to object detection.
2. Upload images (replay frames showing bowler’s front foot near the crease).
3. Label consistently:
   • Class: front_foot (tight bounding box).
   • Keep visible reference to the popping crease for later calibration (not strictly required in labels).
4. Generate dataset in YOLOv8 (Ultralytics) format (train/val split suggested: 70/20/10).
5. Export/download the dataset and fine-tune YOLOv8n in Colab until validation improves.
6. Save the checkpoint with highest validation score as best.pt.

Notes: Small, high-quality labels beat large, noisy datasets. Start with ~30–50 crisp images from your camera style and iterate.

---

How I Ran Inference (Using best.pt)

1. Open the inference notebook notebooks/02_infer_video.ipynb in Google Colab and enable GPU.
2. Upload:
   • The trained best.pt.
   • A short MP4 clip (downloaded from YouTube; trimmed to one delivery).
3. Run the notebook cells in order. The pipeline:
   • Loads the model and the video.
   • Processes frames and draws front-foot overlays.
   • Exports outputs/pred.mp4 as the annotated result.
4. (Optional) Enable Calibration and Rule Decision sections in the notebook to display a LEGAL / NO-BALL verdict (toe vs crease with a small tolerance).

---

Results (Placeholders)

📸

<img width="1650" height="900" alt="image" src="https://github.com/user-attachments/assets/ac582108-08c1-49c3-adfe-82bf17c5dea5" />

---

Requirements

• Google Colab with GPU runtime.
• Short delivery videos (6–12s) where the front foot & crease are visible.
• best.pt (if running inference only) or a small YOLO-format dataset (if you want to train).

---

---

Best Practices

• Use replay-style close-ups; keep clips short and steady.
• For training, prioritize tight, consistent labels over quantity.
• Calibrate once per camera angle for robust LEGAL / NO-BALL decisions.

---

Limitations

• Sensitive to camera angle, blur, and occlusions.
• Performance depends on data quality and representativeness.
• MVP scope is front-foot context; other rules need additional modules.
• For education/training; not an official officiating system.

---

License

MIT License — Use responsibly; respect content rights when sourcing videos.

---

Credits

Ultralytics YOLO, OpenCV, Roboflow, and the open tooling community.