Paper in the making

---

ACE-Trajectories_noTosses 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.

This datasets serves as an enriched version of the original TrackNet Tennis dataset. It cuts the whole rallies included in the TrackNet dataset in single trajectories, based on TrackNets property “Trajectory Pattern”. It supplements the original broadcast TrackNet data with manual ball spin annotations and 2D human pose estimations for the hitting player. It was specifically designed to be used as a validation and test set (split 33% to 66% respectively) for the 2D-to-3D trajectory uplifting models available in the tennisuplifting GitHub repository. This dataset does not contain any images but uses primarily coordinates for indicating the position of the ball, the 16 court keypoints (see image) as well as the 17 COCO Whole Body keypoints (see end of file) of the hitting player on the respective image.

Serves in this dataset do not contain ball tosses. For included ball tosses go to ACE-Trajectories_withTosses.

Original TrackNet Data

The foundation of this dataset is the TrackNet dataset, which consists of broadcast tennis videos and 2D ball tracking data. The original dataset was introduced in this paper and downloaded from Kaggle.

---

Folder Structure

The repository is organized into 517 sequentially numbered folders, ranging from trajectory_0000 to trajectory_0516. Each folder represents a single, continuous tennis trajectory like groundstroke, volley, smash and serves (without the preceding ball toss).

Inside each trajectory_xxxx folder, you will find exactly eight separate .npy files and one info.json file containing the isolated data for that specific sequence.

ACE-Trajectories/
├── Readme.md
├── data/
│   ├── trajectory_0000/
│  │  ├── 2dPoseEstimation.npy
│  │  ├── Mext.npy
│  │  ├── Mint.npy
│  │  ├── hits.npy
│  │  ├── r_img.npy
│  │  ├── spin_class.npy
│  │  ├── start_serve.npy
│  │  ├── times.npy
│  │  └── info.json
│  ├── trajectory_0001/
│  │  ├── 2dPoseEstimation.npy
│  │  ├── Mext.npy
│  │  ├── Mint.npy
│  │  ├── hits.npy
│  │  ├── r_img.npy
│  │  ├── spin_class.npy
│  │  ├── start_serve.npy
│  │  ├── times.npy
│  │  └── info.json
│  │ ...
│  │ ...
│  └── trajectory_0516
│  │  ├── 2dPoseEstimation.npy
│  │  ├── Mext.npy
│  │  ├── Mint.npy
│  │  ├── hits.npy
│  │  ├── r_img.npy
│  │  ├── spin_class.npy
│  │  ├── start_serve.npy
│  │  ├── times.npy
│  │  └── info.json

---

The JSON File

• info.json: This file contains the essential metadata linking this trajectory back to the original TrackNet data. The fields include the specific game and clip the trajectory was cut from, as well as the exact start frame number and end frame number relative to the original TrackNet video.

The NPY Files

The eight numpy arrays store the spatial, temporal, and pose data for the trajectory:

• Mint.npy: A [3 x 3] array representing the reverse-engineered intrinsic camera matrix.
• Mext.npy: A [3 x 4] array representing the reverse-engineered extrinsic camera matrix (containing the rotation matrix and translation vector).
• times.npy: A 1D array of length [T] (where T is the total number of frames in the trajectory) containing the exact timestamps for each frame.
• r_img.npy: A [T x 3] array containing the 2D ball tracking data per frame. The three columns represent the u (horizontal) coordinate, the v (vertical) coordinate, and a visibility class. The visibility class is binary, where 0 means the ball is out of frame/occluded, 1 means visible or hard to spot. The visibility class was directly extracke out of the TrackNet Dataset.
• 2dPoseEstimation.npy: A [17 x 3] array containing the 2D human pose estimation of the hitting player when the racket touches the ball. The rows correspond to the 17 COCO-WholeBody keypoints, and the columns represent the u coordinate, v coordinate, and a model confidence score. For whole serves, this pose is captured at the specific frame where the ball leaves the server's hand (or the first frame if the toss isn't visible).
• spin_class.npy: A 1D array of length [T] repeating the initial spin class for the length of the trajectory. The classes are categorized as 1 (topspin) and 2 (backspin)
• start_serve.npy: Only relevant for serves: A scalar giving the image number referencing the Tracknet dataset that corresponds to the frame where the rackets touches the ball. (relevant fo)
• hits.npy: An array of length [H] providing the time stamps when the ball bounces on the floor. Normally, this H = 1 or H = 0, but if the ball is the final in the rally H can be greater 1.

Download and Usage

This dataset utilizes Git LFS (Large File Storage) for binary .npy files. To ensure that the actual data is downloaded instead of small text pointers, please use the huggingface_hub library.

Installation

pip install huggingface_hub numpy

Dowload the full dataset

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="XSpaceCoderX/ACE-Trajectories_noTosses", 
    repo_type="dataset", 
    local_dir="./data"
)

Loading the data

import numpy as np

# Example: Loading a specific file after download
data = np.load("./data/path/to/file.npy", allow_pickle=True)
print(f"Data shape: {data.shape}")