This repository contains the code and resources for my thesis project, "Graph Neural Networks for event detection in footbal" The project involves building a tool for event spotting in football matches using advanced machine learning techniques including GNNs, Context Aware Loss Function, NetVLAD, and Convolutional Neural Networks to capture both spatial and temporal information.

The goal of this project is to develop a tool that can accurately spot events in football matches by using Graph Neural Networks (GNNs) and other advanced techniques. The model is divided into two main modules: segmentation and spotting. The segmentation module encodes graph representations of video clips into contextual data using GNNs and a context-aware loss function. The spotting module then decodes this contextual information to detect specific events.

The following image illustrates the pipeline of the model, showcasing the segmentation and spotting modules:

The thesis documentation and presentation are included in the thesis_files directory:

• Thesis Document: thesis_files/Thesis_Karol_Ziolo.pdf
• Final Presentation: thesis_files/thesis_presentation.pptx

The project is organized into the following directories and files:

• src: Contains all source code and subfolders
  • data_management: Files for data preprocessing and loading
  • module: Scripts for training models, calibrators, and animation/evaluation classes
  • models: Trained detectors used in the thesis
  • calibrator: Calibration models
  • analysis: Files for generating game animations and visualizing losses or evaluation metrics
  • helpers: Helper functions
  • GPUtraining: Scripts for training models on the GPU lab
  • notebook: Notebooks for analysis and evaluations
  • model.py: Contains all model architectures used
  • Dockerfile: Builds a Docker image for training models on the GPU lab

To install the necessary dependencies, run:

pip install -r requirements.txt

1. Data Management:

   • Preprocess and load data using the scripts in the data_management folder.

2. Training Models:

   • Use the scripts in the GPUtraining folder to train models and calibrators.

3. Animation and Evaluation:

   • Generate game animations and visualize evaluation metrics using scripts in the analysis folder.

4. Running on GPU Lab:

   • Use the GPUtraining scripts and the Dockerfile to set up and train models on a GPU environment.

• Open and run the notebooks in the notebook folder for detailed analysis and evaluations.

To train a model, navigate to the src directory. Fill the paths in the script you plan to run and then run the following command:

Train segmentation model:

python GPUtraining/trainGAT.py

Train spotting model:

python GPUtraining/trainGAT_spotting.py

To generate animations:

python analysis/top_missed_predictions.py -m models/spotting_GAT.pth.tar -a "Shot" -t 5 -o animations/shot