Adit-jain/Soccana_player_ball_detection_v1

⚽ Soccer Object Detection Dataset (25K Subset from 1M+ Images)

---

Index

1. Dataset Overview
2. Folder Structure
3. Dataset Preparation
4. Data Utils
5. Samples

Dataset Overview

This dataset is a curated subset (25,000 images) from a larger soccer vision dataset containing over 1 million images (50+ GB). The data was collected and augmented from multiple open-source sources, including the SoccerNet dataset, video game renders, and publicly available match footage.

It is optimized for object detection tasks, especially focusing on soccer-related entities such as players, referees, and the ball, including various augmentation types like background-only and noisy scenes.

• ✅ 25,000 images (~1.5GB)
• ✅ Annotations for 3 object classes:
  • player
  • referee
  • ball
• ✅ Data format:
  • Ultralytics YOLO format (default)
  • COCO JSON format (included in separate folders)
• ✅ Resolution variety:
  • 160x160
  • 320x320
  • 640x640
  • 1280x1080 (Full HD)
• The dataset includes frames for various scenarios, such as:
  • Occlusions
  • Close up shots
  • Behind the goalpost scenes
  • Camera overlay scenes
  • Low and High angle shots
  • Low resolution shots

• Classes

  Class ID	Label
  0	Player
  1	Referee
  2	Ball

In all, the dataset provides a apt starting point for an all rounder football object detection model.

---

🗂️ Folder Structure

V1/
├── images/
│ ├── train/
│ └── test/
├── labels/ # YOLO TXT labels
│ ├── train/
│ └── test/
├── coco_test_annotations/ # COCO format labels (train.json, val.json)
├── coco_train_annotations/ # COCO format labels (train.json, val.json)
├── data.yaml # Ultralytics YOLOv8-compatible YAML
└── samples/ # Dataset samples

---

Dataset Preparation

Processing Pipeline Architecture

Raw COCO Datasets
        ↓
SAHI Slicing (160/320/640/1280)
        ↓
Image Limit and Filtering
        ↓
Class Name Standardization
        ↓
COCO to YOLO Conversion
        ↓
Final Training Dataset

Raw COCO Datasets:

The following datasets were used for the raw images

1. Soccer Player Tracker (spt_v2)
2. Football Detection Test (tbd_v2)
3. VA Project (v2_temp)
4. Player Detection GKLRL (v12)
5. Football EITPT (v5_temp)
6. Detect Players DGXZ0 (v3)
7. Football Player Detection KUCAB (v7)
8. Football Players Detection 3ZVBC

SAHI slicing

SAHI (Slicing Aided Hyper Inference) is implemented to handle the multi-scale nature of soccer scenes:

Why SAHI for Soccer?

• Crowded Scenes: Penalty area situations with multiple overlapping players
• Scale Variation: Players appear at different sizes based on camera distance
• Small Object Detection: Ball detection in wide-angle shots
• Context Preservation: Maintains spatial relationships through overlapping

slice_sizes = [160, 320, 640, 1280]  # Multiple scale processing
overlap_ratio = 0.2            # 20% overlap between patches

• 160x160 patches: Optimized for small player detection and crowded scenes
• 320x320 patches: Balanced approach for medium-distance shots
• 640x640 patches: Preserves context for tactical analysis and large-scale scenes
• 640x640 patches: For best results in HD context

Image Limit and Filtering

Due to SAHI, the resulting dataset had 1M+ images, and more than 30GB of data. Image filtering was applied from each dataset

# Per-dataset image limits for balanced training
image_limits = {
    "spt_v2": 30, "spt_v2_sahi_160": 30, "spt_v2_sahi_320": 40,
    "tbd_v2": -1, "v2_temp": 300, "v2_temp_sahi_160": 300,
    "v2_temp_sahi_320": 400, "v3": 500, "v3_sahi_160": 500,
    "v3_sahi_320": 1000, "v3_sahi_640": 500, "v5_temp": 500,
    "v7": 500, "v7_sahi_160": 500, "v7_sahi_320": 1000,
    "v7_sahi_640": 500, "v12": 200, "v12_sahi_160": 300,
    "v12_sahi_320": 500, "v12_sahi_640": 300,
}

Class name standardization

Every dataset had different classes, hence three common classes were taken out from each sub dataset

• Player Variants: Maps 'Player', 'Team-A', 'Team-H', 'football player', 'goalkeeper', 'Gardien', 'Joueur' → Class 0
• Ball Variants: Maps 'ball', 'Ball', 'Ballon', 'football' → Class 1
• Referee Variants: Maps 'referee', 'Referee', 'Arbitre' → Class 2

COCO to YOLO

the final COCO format dataset was converted to YOLO format fro ultralytics pipeline. Both the formats can be found in the zip file.

---

Data Utils

Processing Scripts Location

All dataset processing utilities are available in the Data_utils directory:

🔗 Repository Link: https://github.com/Adit-jain/Soccer_Analysis/tree/main/Data_utils

Key Utilities

External_Detections/

• slice_images.py: SAHI-based multi-scale slicing
• merge_datasets.py: Multi-dataset integration with class mapping
• coco_to_yolo.py: Format conversion with coordinate normalization
• create_data_yaml.py: YOLO training configuration generation
• visualize_coco_dataset.py: Quality control and visualization

SoccerNet_Detections/

• get_soccernet_data.py: SoccerNet dataset downloading
• data_preprocessing.py: MOT to YOLO conversion pipeline

---

Samples