🏏 Cric-360

Cricket Broadcast Ground Dataset — Version 1.0

🌍 The world's first large-scale, AI-cleaned cricket broadcast ground image dataset for computer vision research.

---

📖 Overview

Cric-360 v1.0 is a carefully curated collection of 3,558 broadcast-quality cricket ground frames extracted from real live match footage captured across ~20 international stadiums from Asia, Africa, SENA (South Africa, England, New Zealand, Australia), and beyond.

This is the initial release of an ongoing project. We plan to grow this dataset significantly over time with more venues, annotations, and modalities.

Unlike existing cricket datasets focused only on action recognition or ball tracking, Cric-360 targets the complete visual understanding of cricket broadcast scenes — from ground segmentation to augmented reality overlays.

All images have been processed through a custom AI cleaning pipeline (LaMa inpainting) to remove broadcast overlays, providing clean, research-ready frames.

---

🌍 Dataset Diversity

This is what makes Cric-360 stand out — exceptional visual diversity across every dimension:

🏟️ Venue Coverage (~20 International Stadiums)

Region	Coverage
🌏 Asia	Pakistan, India, Sri Lanka, Bangladesh, UAE stadiums
🌍 Africa	South Africa venues
🏴󠁧󠁢󠁥󠁮󠁧󠁿 England	English grounds
🇦🇺 Australia / NZ	SENA region stadiums
🌐 Neutral venues	International hosted matches

No single-venue bias — the model trained on this data generalises across diverse cricket ground layouts and designs.

☀️🌙 Lighting Conditions

Condition	Description
☀️ Day matches	Full sunlight, harsh shadows, bright outfield
🌙 Day-Night matches	Transitional lighting, mixed floodlight + sunlight
💡 Night matches	Full floodlight conditions, deep shadows
🌥️ Overcast	Flat diffuse lighting, no hard shadows

📷 Camera Angles & Frame Types

Frame Type	Description
🔭 Wide shots	Full ground view — boundary to boundary
📹 Medium shots	Mid-pitch, wicket-area focused
🔍 Close-up shots	Tight on pitch, crease, batting end
🎥 Side-on angles	Traditional broadcast side camera
🏔️ High cameras	Overhead / spider-cam perspectives
📐 Low angles	Ground-level and near-boundary views

🌑 Shadow & Occlusion Diversity

Shadow Type	Present
Player shadows on pitch	✅
Stadium / stand shadows	✅
Floodlight pole shadows	✅
Partial ground occlusion	✅
Carpet / pitch shadows	✅

---

📊 Dataset at a Glance

📌 Property	📋 Value
Total Images	3,558
Primary Resolution	1920 × 990–1080 px
Format	JPEG
Stadiums	~20 international venues
License	Apache 2.0
Cleaning Method	LaMa AI inpainting + mirror-pad
Split Strategy	Random shuffle · seed = 42
Train / Val / Test	2,490 / 533 / 535
Version	1.0 (initial release)

🏆 Quality Breakdown

Quality Tier	Count	%
🟢 HD (≥1920×1080)	424	11.9%
🔵 HD-Ready (≥1280×720)	3,072	86.3%
🟡 SD (≥854×480)	61	1.7%
🔴 Low	1	0.0%
Total	3,558	100%

---

🎯 Applications & Use Cases

Cric-360 is designed to be a multi-purpose foundation dataset for cricket broadcast computer vision. Here is the full spectrum of what it enables:

🔬 Segmentation Tasks

Task	How Cric-360 Helps
Ground Segmentation	Segment pitch, outfield, boundary rope, grass, stands
Player Segmentation	Isolate players against diverse ground textures
Shadow Segmentation	Diverse natural + artificial shadows for robust models
Pitch / Carpet Detection	Detect drop-in pitch surfaces vs. natural turf

🎯 Detection & Tracking

Task	How Cric-360 Helps
Ball Tracking	Ground-plane context for trajectory reconstruction
Player Tracking	Multi-camera perspective diversity for robust detectors
Boundary / Rope Detection	Identify playing field limits from broadcast views
Ad Board Detection	Locate physical and virtual advertisement boards

🗺️ Geometry & Calibration

Task	How Cric-360 Helps
Homography Estimation	Map broadcast frame → top-down pitch diagram
Camera Calibration	Intrinsic/extrinsic estimation from broadcast footage
Camera Pose Estimation	Multi-angle diversity for pose recovery algorithms
3D Ground Reconstruction	Depth and geometry from diverse camera positions

🎮 Augmented Reality & Broadcast

Task	How Cric-360 Helps
Virtual Advertisement Insertion	Clean ground plane for digital overlay placement
3D Logo Placement	Ground surface estimation for AR brand overlays
Broadcast Enhancement	Scorebar / overlay repositioning pipelines
Virtual Pitch Replacement	Replace natural pitch with synthetic for broadcast

🤖 Scene Understanding

Task	How Cric-360 Helps
Depth Estimation	Monocular depth on complex stadium scenes
Scene Classification	Day/night, venue type, match phase classification
Lighting Estimation	Shadow diversity for outdoor scene illumination models
Domain Adaptation	Bridge synthetic → real cricket scene distribution

---

✨ What Makes Cric-360 Unique

Feature	Details
🥇 World's First	First publicly released cricket ground image dataset
🌍 ~20 Stadiums	Asia, Africa, SENA — no venue bias
☀️🌙 Day + Night	All lighting conditions covered
📷 All Angles	Wide, medium, close; side-on, high, low cameras
🌑 Shadow Diversity	Player, stadium, floodlight, and pitch shadows
🤖 AI-Cleaned	LaMa inpainting removes all broadcast overlays
🎯 Carpet Diversity	Drop-in pitch and natural turf frames
🔀 Bias-Free Splits	Randomly shuffled with seed=42 before split
📈 Living Dataset	v1.0 — actively maintained and expanded

---

🗂️ Dataset Structure

Cric-360/
├── data/
│   ├── train/          ← 2,490 images  (70%)
│   ├── val/            ← 533  images   (15%)
│   └── test/           ← 535  images   (15%)
├── metadata.csv        ← Per-image metadata (all 3,558 rows)
└── README.md

📄 metadata.csv — Column Reference

Column	Type	Description
file_name	string	Path · e.g. data/train/frame_001.jpg
split	string	train / val / test
width	int	Image width in pixels
height	int	Image height in pixels
resolution	string	{width}x{height}
quality	string	HD / HD-Ready / SD / Low
venue	string	Stadium name (where identifiable)
has_carpet	string	yes / no / unknown
image_mode	string	PIL colour mode (RGB, etc.)
format	string	JPEG / PNG

---

🚀 Quick Start

Load with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("sarimshahzad/Cric-360")
print(ds)
# DatasetDict({
#     train: Dataset({num_rows: 2490})
#     val:   Dataset({num_rows: 533})
#     test:  Dataset({num_rows: 535})
# })

# Access a sample
sample = ds["train"][0]
image  = sample["image"]           # PIL Image
print(image.size)                  # (1920, 990)
print(sample["quality"])           # 'HD-Ready'

# Filter by quality
hd = ds["train"].filter(lambda x: x["quality"] == "HD")

CLI Download

pip install huggingface_hub
huggingface-cli download sarimshahzad/Cric-360 --repo-type dataset --local-dir ./Cric-360

Git Clone

git lfs install
git clone https://huggingface.co/datasets/sarimshahzad/Cric-360

---

🧹 Dataset Cleaning Pipeline

Raw Broadcast Frame
       │
       ▼
┌──────────────────────────────┐
│  Step 1: TV Logo Removal     │  LaMa inpainting on corner regions
└──────────────────────────────┘
       │
       ▼
┌──────────────────────────────┐
│  Step 2: Scorebar Removal    │  Mirror-pad: reflects ground pixels
│  (bottom strip / overlays)   │  giving LaMa photorealistic context
└──────────────────────────────┘
       │
       ▼
┌──────────────────────────────┐
│  Step 3: QA Verification     │  Manual spot-check on inpainted zones
└──────────────────────────────┘
       │
       ▼
  Clean Broadcast Frame ✅

This is the first dataset to publicly document AI-based overlay removal as part of a sports broadcast dataset curation pipeline.

---

🔀 Bias Prevention: Split Strategy

To ensure no temporal ordering or venue concentration bias between splits:

1. All 3,558 images randomly shuffled (random.seed(42))
2. Split sequentially after shuffle: 70% train · 15% val · 15% test
3. This prevents any single match or venue from dominating a split

---

🔭 Version History & Roadmap

v1.0 — Current Release (March 2025)

• ✅ 3,558 AI-cleaned broadcast frames
• ✅ ~20 international stadiums
• ✅ Day, day-night, and night conditions
• ✅ All camera angles and frame distances
• ✅ Full train/val/test split with metadata

🚀 Planned — v2.0 and Beyond

• 🔲 Semantic segmentation masks (ground, pitch, outfield, stands)
• 🔲 Player bounding box annotations
• 🔲 Homography ground-truth matrices (for registered pairs)
• 🔲 10,000+ image target
• 🔲 Multi-match temporal sequences
• 🔲 Night-only subset with floodlight metadata
• 🔲 Carpet / no-carpet balanced labels verified by hand

Cric-360 is a living dataset. Star ⭐ the repo and watch for updates.

---

🔗 Related Datasets

Dataset	Domain	Notes
SoccerNet	Soccer	Broadcast video, action spotting
WorldCup2014	Soccer	Homography / field registration
CricShot10	Cricket	Shot classification only
DeepSport	Basketball	Court segmentation

⚠️ No comparable cricket ground dataset exists publicly. Cric-360 is the first.

---

📜 Citation

@dataset{cric360_2025,
  author    = {Shahzad, Sarim},
  title     = {Cric-360: A Cricket Broadcast Ground Dataset for Computer Vision (v1.0)},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/sarimshahzad/Cric-360},
  note      = {Version 1.0 — initial release}
}

---

📬 Contact & Contributions

🤗 HuggingFace	@sarimshahzad
🐛 Issues / Requests	Use the Community tab above
🤝 Collaboration	Open to annotation partners and research collaborators

---

Apache 2.0 — Free for academic and commercial use with attribution.

This is v1.0 — the beginning. We're just getting started. 🏏

Made with ❤️ for the global computer vision research community.