Silent Impact

This is the official repository of the paper
"Silent Impact: Tracking Tennis Shots from the Passive Arm"

https://doi.org/10.1145/3654777.3676403

Dataset

The dataset is categorized into two main components: Classification and Detection. Both categories contain data from IMUs on the dominant and passive arms of participants.

Download

The dataset can be downloaded here. Place the 4 pkl files in the data folder so that it looks like this:

Silent-Impact/
    data/
        Classification_Dominant.pkl
        Classification_Passive.pkl
        Detection_Dominant.pkl
        Detection_Passive.pkl
        Participant_Information.json
    README.md
    ...

Classification Data

Files:

• Dominant Arm: Classification_Dominant.pkl
• Passive Arm: Classification_Passive.pkl

Description & Data Structure:

The classification array contains 6,000 shots across 6 different shot types, performed by 20 participants. Each entry in this array represents a sequence of sensor readings across 6 channels and spans 180 frames. At the beginning of each entry, you'll find the shot type label (ranging from 0 to 5) and the participant's ID.

• Array Shape: [number_of_items, meta_data_size + window_length, sensor_channels]

  • number_of_items: 6,000
    (Calculated as: 20 participants × 6 shot types × 50 shot instances)

  • meta_data_size: 1

    • Index 0: Shot type label
      • Labels:
        0 - Forehand Stroke
        1 - Backhand Stroke
        2 - Forehand Volley
        3 - Backhand Volley
        4 - Smash
        5 - Serve
    • Index 1: Participant ID

    For instance, classification_array[:, 0, 0] returns the shot type of each entry.

  • window_length: 180

  • sensor_channels: 6

    • 3 channels: X/Y/Z linear acceleration
    • 3 channels: X/Y/Z angular velocity

Detection Data

Files:

• Dominant Arm: Detection_Dominant.pkl
• Passive Arm: Detection_Passive.pkl

Description & Data Structure:

The detection array provides sensor readings from tennis sessions, including rallies and casual matches, from 10 participants. Every entry in this array features a 6-channel sensor reading sequence, accompanied by a shot/no-shot label for each frame. As the sessions vary in length, the array's length matches the longest session. Each entry starts with the length of its respective sequence and the ID of the associated participant.

• Array Shape: [number_of_items, meta_data_size + longest_sequence_length, sensor_channels + frame_label_size]

  • meta_data_size: 1

    • Index 0: Sequence length
    • Index 1: Participant ID

    For instance, detection_array[:, 0, 0] returns the length of each sequence.

  • sensor_channels: 6

    • 3 channels: X/Y/Z linear acceleration
    • 3 channels: X/Y/Z angular velocity

  • frame_label_size: 1

    • 0: No shot
    • 1: Shot

Subject information

Demographics and tennis characteristics of each subject is provided in the Participant_Information.json file.

Code

Tested with python==3.7.16, pytorch==1.13.0, cudatoolkit=11.6

Dependencies

conda create -n silentimpact python=3.7
conda activate silentimpact
conda install pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0 pytorch-cuda=11.6 -c pytorch -c nvidia
pip install -r requirements.txt

Classification training

python train_classification.py --arm p --name test --split 1 --gpu 0

Detection training

python train_detection.py --arm p --name test --split 1 --gpu 0

Citation

If you use the dataset or code provided in this work, please cite us:

TBD