G.O.A.T Vision LIVE DEMO https://goatvision-soccer.streamlit.app
ITAI-1378 Midterm – G.O.A.T Vision: Game Outcome Action Tracker Project Name
G.O.A.T Vision (Game Outcome Action Tracker) is a computer vision application that classifies soccer plays from images using a deep learning model (ResNet50). Users can upload an image and receive a prediction along with confidence scores.
Course: ITAI 1378 – Computer Vision Student: Samira Gevara
Project Tier
Tier 2 – Custom Dataset + Transfer Learning
This project builds a custom image classification system using transfer learning with a pretrained deep learning model.
Problem Statement
Analyzing soccer events from visual footage is often a manual and time-consuming process for analysts, coaches, and content creators. Automatically identifying key moments such as free kicks, penalty kicks, corner kicks, shots, and yellow cards can significantly improve the efficiency of highlight creation and match analysis.
This project explores how computer vision can classify soccer event images into specific play types.
Solution Overview
G.O.A.T Vision is a computer vision model that classifies soccer event images into five categories:
Corner Kick
Free Kick
Penalty Kick
Shot on Goal
Yellow Card
Displays prediction confidence
Visual probability breakdown
The system uses transfer learning with a pretrained ResNet50 model implemented in PyTorch. Images are preprocessed and passed through the neural network to produce a predicted soccer event class.
Technical Approach
Technique: Image Classification
Model: ResNet50 (pretrained on ImageNet)
Frameworks and Tools:
PyTorch
torchvision
scikit-learn
Google Drive (model hosting via gdown)
Streamlit
matplotlib
Google Colab (GPU)
Transfer learning allows the model to leverage features learned from millions of images while adapting to a smaller soccer-specific dataset.
Dataset Plan
Dataset Size: 146 curated images
Classes:
corner_kick
free_kick
penalty_kick
shot
yellow_card
Images were collected from publicly available soccer imagery sources and manually cleaned to remove irrelevant or ambiguous images such as cartoons, logos, or unclear events.
Deployment Notes
Due to GitHub file size limits, the trained model is hosted externally on Google Drive. The app automatically downloads the model at runtime using gdown.
Data Augmentation Used:
Horizontal flip
Color jitter
Image resizing
Image normalization
These augmentations help improve generalization despite the small dataset.
Model Evaluation
Primary Metric: Validation Accuracy
Target Accuracy: ≥ 70% Achieved Accuracy: 82.14%
Secondary Evaluation:
Confusion Matrix analysis
Prediction visualization on validation images
Key observations from evaluation:
Yellow card detection achieved perfect classification in validation samples.
Some confusion occurred between shots, penalty kicks, and corner kicks due to similar goal-area visual contexts.
System Pipeline Input Soccer Image ↓ Image Preprocessing (resize, normalization, augmentation) ↓ ResNet50 Feature Extraction (transfer learning) ↓ Fully Connected Classifier ↓ Predicted Event Label (Corner Kick | Free Kick | Penalty Kick | Shot | Yellow Card)
Week-by-Week Plan Week Task Milestone Week 11 Dataset collection and cleaning Dataset prepared Week 12 Model setup and preprocessing Training pipeline ready Week 13 Train model and test predictions Model learning Week 14 Evaluate results and confusion matrix Performance analysis Week 15 Prepare slides and documentation Presentation ready Week 16 Present project Final submission (May 8th)
Resources Needed Resource Details Compute Google Colab GPU Frameworks PyTorch, torchvision Libraries sklearn, matplotlib Estimated Cost $0
Risks and Mitigation Risk Probability Mitigation Small dataset size Medium Data augmentation and transfer learning Ambiguous images Medium Manual dataset cleaning Class confusion Low Model fine-tuning and confusion matrix analysis
AI Usage Log
AI tools such as ChatGPT were used to assist with:
project planning
code debugging
PyTorch implementation guidance
documentation drafting
All code execution, dataset preparation, and model experimentation were performed by the student (Samira Gevara)
Repository Structure project-name/ README.md requirements.txt notebooks/ 01_exploration.ipynb data/ README.md docs/ proposal.pdf
Future Improvements
Potential improvements for future development include:
expanding the dataset to thousands of labeled soccer event images
training with video frame sequences instead of single images
implementing object detection to localize players and referees
building a real-time match event detection system
