Predictive Pre-Training of Tennis Shot Embeddings

Introduction

The goal of the project is to perform time series prediction of player poses and ball trajectory in tennis matches. In order to limit computational complexity, a feature extraction pipeline will be built using open source models for object detection and pose estimation, to convert dense video into a sparse 3D representation of the point. These extracted features will serve to build a time series prediction model based on graph neural networks and recursive neural networks. The ability to predict the movement of players and ball in a tennis point is projected to unlock many downstream applications, from analytics to coaching.

The authors of this project are:

• Tawab Safi (as7092)
• George Tamer (gyt2107)
• Flor Sanders (fps2116)

Repository Structure

The repository is structured as follows.

• src: The source directory contains the project's implementation code, divided into four subparts: data, model, train and eval.
• data: This directory is where the data set should be stored, it can be downloaded from Google Drive.
• models: Our pre-trained model weights are saved in this directory. Additionally, the model weights needed for pre-processing should be saved here.

Finally, the root directory of this repository also contains the project slides and report.

Setup

Clone the repository to make a local copy and change directory.

git clone [email protected]:FlorSanders/Tennis_Shot_Prediction.git
cd adl_ai_tennis_coach

The dependencies to run the code are included as an environment.yml file.
They can be installed using conda by executing the following commands.

conda env create -f environment.yml
conda activate teco

Separate install instructions are required for the data pre-processing.
These are provided in the README of the src/data directory.