This Python project utilises neural networks to train a model that can predict the location of balls in sports photographs, specifically scenarios where the ball is not visible.

The goal? To effectively predict the balls position in images where it has been digitally removed.

Follow these steps to create and activate a virtual environment:

1. python -m venv venv
2. source venv/bin/activate
3. pip install -r requirements.txt

Collect Images (collect_images.py): Use to fetch 'spot the ball' images from a specified source. Images are stored temporarily on the local disk in two directories:

• images_raw/{id}.png
• images_training/{id}--{x_coord}-{y_coord}.png for training images, where filenames include the correct ball coordinates.

Generate a Trained Model (trainer.py): Execute to train your model using the collected images. The script outputs a model file that will be used for making predictions.

With a trained model, you can predict the ball's location in new images.

Training Predictor (training_predictor.py): This script processes images from the images_training directory, which contain known ball positions. It overlays both the actual and predicted ball locations on the images, marking them with crosses of different colors. It's particularly useful for evaluating the model's accuracy.

General Predictor (predictor.py): Use this script to predict the ball's location in images without known coordinates. Predictions are marked with a cross on the output images.

This neural network serves as a starting point, there is ample room for further optimisation. Enhancing the models training process and predictions is key to improving the accuracy.