README.md

---
license: cc-by-nc-4.0
task_categories:
- video-classification
- time-series-forecasting
- object-detection
language:
- en
- zh
tags:
- sports
- physics
- table-tennis
- trajectory-prediction
- ping-pong
size_categories:
- 1K<n<10K
extra_gated_fields:
  First Name: text
  Last Name: text
  Affiliation: text
  Research Purpose: text
  I agree to use this dataset strictly for academic purposes and will not attempt to identify individuals: checkbox
---
## 🏓 BNDSTT Dataset Description

### 1. Dataset Overview & Motivation
Existing general sports datasets (e.g., Kinetic 400) contain limited table tennis footage primarily designed for action classification. Other specific datasets, such as Baidu's P2Anet, label service actions but lack critical data on **ball spin** and **landing point**.

To address these limitations and support the development of AI systems that learn progressively—from coaching demonstrations to amateur play, and finally to professional matches—we constructed the **BNDSTT (Beijing National Day School Table Tennis)** dataset.

### 2. Data Composition
The dataset currently consists of **997** table tennis service videos:
* **20% (200 videos):** High-quality instructional service videos curated from Bilibili.
* **80% (797 videos):** Real-world training and match footage featuring the author and other school team members.

### 3. Preprocessing & Construction
To ensure model focus and accuracy, the data underwent specific preprocessing:
#### **Subject Extraction (YOLO)**
We utilized a YOLO model to detect individuals in the frame. Using a "center-focused and largest-area" principle, the serving player was identified and cropped to minimize background noise.

<div align="center">
  <img src="./preprocessing_yolo.png" width="80%" />
  <p><em>Figure 1: YOLO cropping centered on the serving player.</em></p>
</div>

#### **Trajectory Truncation**
For the landing point prediction task, videos were truncated to the moment the ball crosses the net. The goal is to predict the final landing spot based solely on the trajectory *before* the net.

<div align="center">
  <img src="./preprocessing_trajectory.png" width="80%" />
  <p><em>Figure 2: Video truncation at the net-crossing moment.</em></p>
</div>

---

## 📊 Dataset Statistics

### Spin Distribution (Task 1)
The dataset covers 9 distinct spin types:

| Spin Type | Count |
| :--- | :--- |
| **Backspin** | 181 |
| **Topspin** | 84 |
| **Left Sidespin** | 57 |
| **Right Sidespin** | 34 |
| **Left-Side Topspin** | 130 |
| **Left-Side Backspin** | 134 |
| **Right-Side Topspin** | 168 |
| **Right-Side Backspin** | 204 |
| **No Spin ** | 5 |

### Landing Point Distribution (Task 2)
Landing positions are categorized into 6 zones:

| Landing Position | Count |
| :--- | :--- |
| **Middle Short** | 274 |
| **Left Long** | 274 |
| **Middle Long** | 246 |
| **Left Short** | 97 |
| **Right Long** | 57 |
| **Right Short** | 49 |

---

## 📝 Annotation Format

The annotation file follows a simple tabular structure:

```text
<Filename> <Spin Code> <Landing Code>
```

**Example:**
`TableTennis_00015_RightSideTop_LeftShort.mp4 6 100`

### Label Encoding Reference

**Spin Codes (0-8):**
* `0`: Topspin
* `1`: Backspin
* `2`: Left Sidespin
* `3`: Right Sidespin
* `4`: Left-Side Topspin
* `5`: Left-Side Backspin
* `6`: Right-Side Topspin
* `7`: Right-Side Backspin
* `8`: No Spin

**Landing Codes (100-105):**
* `100`: Left Short
* `101`: Right Short
* `102`: Left Long
* `103`: Right Long
* `104`: Middle Short
* `105`: Middle Long

## 📜 License & Ethical Usage

This dataset is released under the **Creative Commons Attribution-NonCommercial 4.0 International (CC-BY-NC-4.0)** license.

### 🚫 Non-Commercial Use Only
By downloading or accessing this dataset, you agree that:
1. The data will be used **solely for academic research purposes**.
2. Any commercial use, including but not limited to training commercial AI models or use in for-profit products, is **strictly prohibited**.

### 🔒 Ethical & Privacy Notice
This dataset contains video data recorded in a controlled sports environment and was **collaboratively constructed by a team of high school researchers**.

* **Privacy & Consent:** To strictly respect the privacy preferences of all participants, data from students who did not wish to appear publicly has been excluded. **This repository solely contains data from contributors who provided explicit consent for public release.**
* **Facial Recognition Ban:** Users must agree **not** to attempt to identify individuals in the videos or use the data for facial recognition tasks.

---

## 🖊️ Citation

If you use this dataset in your research, please cite it as follows:

```bibtex
@misc{li2025tabletennis,
  author = {Li, Zhanran},
  title = {BNDSTT: Table Tennis Serve Trajectory Dataset},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/J0LE/BNDSTT}},
  note = {Licensed under CC-BY-NC-4.0}
}