Official Repository for KDD'25 (Dataset and Benchmark Track) Paper "A Multimodal Dataset with Targeted Suggestions for LLM-based Virtual Coach in Table Tennis Training "
T3Set (Table Tennis Training) is a multimodal dataset with aligned video-sensor-text data in table tennis training, designed for LLM-based virtual coach research.
- Temporal alignment between sensor data (IMU), multi-angle video, and professional coaching text
- High-quality targeted suggestions following a predefined suggestion taxonomy
| Dimension | Details |
|---|---|
| Participants | 32 amateur players |
| Training Rounds | 380 multi-ball training rounds |
| Strokes | 8,655 labeled strokes |
| Technique | 7 common techniques (topspin, block, etc.) |
| Targeted Suggestions | 8,395 coach suggestions |
| Modal Data | - Video: 1080p@60fps, two-camera - Sensor: 9-axis IMU (100Hz sampling) - Audio: coaches' audio and text |
The dataset is publicly available on Zenodo:
# Access link
https://zenodo.org/records/15516144
# Permanent DOI link
https://doi.org/10.5281/zenodo.15516143 - All participants provided informed consent for open-source usage.
- Experimental procedures approved by the laboratory ethics review committee.
The paper is available on ACM DL:
# Access link
https://dl.acm.org/doi/10.1145/3711896.3737407T3Set/
├─ models/ # A simple model to validate the usage of dataset
│ ├─ src/ # src code
│ ├─ weights/ # Pre-trained weights
│ ├─ requirements.txt # required packages
│ └─ README.md # usage instructions
├─ scripts/ # Data processing & evaluation scripts
│ ├─ data_scripts/ # scripts for building dataset (stroke detection, data alignment, text preprocessing)
│ └─ eval_scripts/ # Benchmark testing script
├─ README.md # Project overview
└─ LICENSE # License information
If you find this dataset useful, please cite our paper:
@inproceedings{ma2025t3set,
author = {Ma, Ji and Wu, Jiale and Wang, Haoyu and Zhang, Yanze and Xie, Xiao and Zhou, Zheng and Zhang, Hui and Wang, Jiachen and Wu, Yingcai},
title = {T3Set: A Multimodal Dataset with Targeted Suggestions for LLM-based Virtual Coach in Table Tennis Training},
year = {2025},
publisher = {Association for Computing Machinery},
address = {New York, NY, USA},
doi = {10.1145/3711896.3737407},
booktitle = {Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2},
pages = {5686–5697},
numpages = {12},
location = {Toronto ON, Canada},
series = {KDD '25}
}Thanks to all players and coaches involved in data collection!
For questions or collaborations, contact:
- Ji Ma: zjumaji@zju.edu.cn