XSpaceCoderX/AD-Rallies

Paper in the making

---

AD-Trajectories Dataset

This dataset was created for the Master's thesis "From Broadcast to 3D: A Deep Learning Approach for Tennis Trajectory and Spin Estimation" by Alexandra Göppert at the University Augsburg, Chair of Machine Learning and Computer Vision. The AD-Rallies dataset is a large-scale synthetic dataset generated using the MuJoCo physics engine. It was built to bridge the synthetic-to-real gap by providing highly accurate physical models of aerodynamic forces, such as the Magnus effect, and complex ball-court interactions.

---

Dataset Overview

The dataset comprises of approxemately 3.2 Million synthetic tennis rallies. The rallies start with a ball toss, a serve and then up to 4 further basic strokes like (groundstroke, volley, lob, short and smash) can be added. All physical kinematics, including the 3D positions, linear velocities, and angular velocities (spin), are captured at a high resolution of 500 frames per second (fps), corresponding to a time step of 0.002 seconds.

The dataset is saved as a .tar file becuase it compromises out of 3.2 Million .npz files. The tar file has a size of 87 GB. The npz file includes the position, velocity and angular velocity of the ball through the whole rally.

The name if the npz files is like follows: toss_xxxxx_branch_yyy.npz or toss_xxxxx_branch_yyy_deadend.npz

Where xxxxx is the number of one of the 20000 tosses that were initially simulated and start the rally. Combined with a server this creates the stem for the rallies. For each stem rally upto 4 returns are added. These are numbered in increasing order as branch_yyy. The number of max rallies emerging out of one toss-serve combination is 160. If there is no feasable return found the rally is no longer stritched together (even though it has less than a total of 6 shots in the rally). This rally is then marked with "_deadend".

---

Data Structures per Trajectory

Inside each npz, you will find exactly seven .npy files. These numpy arrays store the spatial, temporal, and camera data for that specific sequence:

positions.npy: The 3D position of the ball (x, y, z) throughout the rally, recorded at a resolution of 0.002s.
velocities.npy: The linear velocity of the ball relative to the world coordinate system, recorded at a resolution of 0.002s.
rotations.npy: The angular velocity (spin) of the ball in all 3 directions, recorded at a resolution of 0.002s.

The position and velocity is defined in relation to the 3D world coordinate system, which is defined like follows:

The ball spin (rotations.npy) is defined in relation to the ball's local coordinate system. The direction of which is defined as follows:

---

Download the Dataset

You can download the specific tar file using the hf_hub_download function. This is more efficient than cloning the entire repository if you only need the archive. Python

from huggingface_hub import hf_hub_download

REPO_ID = "XSpaceCoderX/AD-Rallies"
FILENAME = "data.tar"

print(f"Downloading {FILENAME}...")

local_path = hf_hub_download(
    repo_id=REPO_ID,
    filename=FILENAME,
    repo_type="dataset"
)

print(f"File downloaded to: {local_path}")

Unpack the Dataset

Once downloaded, you can extract the contents using Python's built-in tarfile module or a system command. Option A: Using Python (Cross-platform)

This is the recommended way to ensure compatibility across Windows, macOS, and Linux.

import tarfile
import os

def extract_tar(file_path, extract_path="."):
    print(f"Extracting {file_path}...")
    with tarfile.open(file_path, "r") as tar:
        tar.extractall(path=extract_path)
    print("Extraction complete!")

Note: Make sure you have at least 180GB of free disk space (87 GB for the archive + 100 GB for the extracted contents).