README.md

---
license: mit
language:
  - en
pretty_name: RacketVision Dataset
size_categories:
  - 10K<n<100K
task_categories:
  - object-detection
  - video-classification
tags:
  - sports-analytics
  - computer-vision
  - object-tracking
  - trajectory-prediction
  - ball-tracking
  - racket-pose-estimation
  - badminton
  - table-tennis
  - tennis
  - racket-sports
---

# RacketVision Dataset

[![Arxiv](https://img.shields.io/badge/ArXiv-2511.17045-B31B1B.svg)](https://arxiv.org/abs/2511.17045)
[![AAAI](https://img.shields.io/badge/AAAI_2026-Oral-blue.svg)](https://aaai.org/)
[![GitHub](https://img.shields.io/badge/GitHub-Code-black?logo=github)](https://github.com/OrcustD/RacketVision/)
[![Hugging Face Dataset](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Dataset-blue)](https://huggingface.co/datasets/linfeng302/RacketVision)
[![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-yellow)](https://huggingface.co/linfeng302/RacketVision-Models)

**RacketVision** is a large-scale, multi-sport dataset and benchmark for advancing computer vision in sports analytics, covering **badminton**, **table tennis**, and **tennis**. It is the first dataset to provide large-scale, fine-grained annotations for racket pose alongside traditional ball positions, enabling research into complex human-object interactions. The benchmark tackles three interconnected tasks: fine-grained **ball tracking**, articulated **racket pose estimation**, and predictive ball **trajectory forecasting**.

![Teaser](https://raw.githubusercontent.com/OrcustD/RacketVision/main/assets/teaser.jpg)

## Using this Hub repository

This dataset is distributed as **static files** (videos, CSV, JSON, PKL). Download it with the Hugging Face CLI, then follow the [project README](https://github.com/OrcustD/RacketVision/blob/main/README.md) for environment setup and training:

```bash
# Official code layout (clone https://github.com/OrcustD/RacketVision ): from repo root
hf download linfeng302/RacketVision --repo-type dataset --local-dir source/data

# Stand-alone data folder only (you must point module configs or --data_root to this directory)
hf download linfeng302/RacketVision --repo-type dataset --local-dir data
```

The in-browser Dataset Viewer may not fully load all assets: COCO detection and pose JSON files use different annotation schemas, so they are not merged into a single `datasets`-style table. Use the files on disk as documented below.

## Directory Layout

```
data/
├── annotations/
│   └── dataset_info.json          # Global dataset metadata (clip list, splits)
│
├── info/                          # COCO-format annotations for RacketPose
│   ├── train_det_coco.json        # Detection: bbox annotations (train split)
│   ├── val_det_coco.json
│   ├── test_det_coco.json
│   ├── train_pose_coco.json       # Pose: keypoint annotations (train split)
│   ├── val_pose_coco.json
│   └── test_pose_coco.json
│
├── <sport>/                       # badminton / tabletennis / tennis
│   ├── videos/
│   │   └── <match>_<rally>.mp4    # Raw video clips
│   ├── all/
│   │   └── <match>/
│   │       ├── csv/<rally>_ball.csv        # Ball ground truth annotations
│   │       └── racket/<rally>/*.json       # Racket ground truth annotations
│   ├── interp_ball/               # Interpolated ball trajectories (for rebuilding TrajPred data)
│   ├── merged_racket/             # Merged racket predictions (for rebuilding TrajPred data)
│   └── info/
│       ├── metainfo.json          # Sport-specific metadata
│       ├── train.json             # [[match_id, rally_id], ...] for training
│       ├── val.json               # Validation split
│       └── test.json              # Test split
│
└── data_traj/                     # Pre-built trajectory prediction datasets
    ├── ball_racket_<sport>_h20_f5.pkl    # Short-horizon: 20 history → 5 future
    └── ball_racket_<sport>_h80_f20.pkl   # Long-horizon: 80 history → 20 future
```

**Local preprocessing (required for BallTrack):** after download, generate per-match `frame/<rally>/` (JPG frames) and `median.npz` from the videos using `DataPreprocess/extract_frames.py` and `DataPreprocess/create_median.py`. These are omitted from the Hub release to save space; see the [project README](https://github.com/OrcustD/RacketVision/blob/main/README.md).

## Data Formats

### Ball Annotations (`csv/<rally>_ball.csv`)

| Column | Type | Description |
|--------|------|-------------|
| Frame | int | 0-indexed frame number |
| X | int | Ball center X in pixels (1920×1080) |
| Y | int | Ball center Y in pixels |
| Visibility | int | 1 = visible, 0 = not visible |

### Racket Annotations (`racket/<rally>/<frame_id>.json`)

Per-frame JSON with a list of racket instances, each containing:

```json
{
  "category": "badminton_racket",
  "bbox_xywh": [x, y, w, h],
  "keypoints": [[x1, y1, vis], [x2, y2, vis], ...]
}
```

**5 keypoints** are defined: `top`, `bottom`, `handle`, `left`, `right`.

### COCO Annotations (`info/*_coco.json`)

Standard COCO format used by RacketPose for training/evaluation:

- **Detection** (`*_det_coco.json`): 3 categories — `badminton_racket`, `tabletennis_racket`, `tennis_racket`.
- **Pose** (`*_pose_coco.json`): 1 category (`racket`) with 5 keypoints.

### Trajectory PKL (`data_traj/*.pkl`)

Pickle files containing pre-processed sliding-window samples. Each PKL has:

```python
{
    'train_samples': [...],   # List of sample dicts
    'test_samples': [...],
    'train_dataset': ...,     # Legacy Dataset objects
    'test_dataset': ...,
    'metadata': {
        'history_len': 80,
        'future_len': 20,
        'sports': ['badminton'],
        'total_samples': N,
        'train_size': ...,
        'test_size': ...
    }
}
```

Each sample dict:

```python
{
    'history': np.array(shape=(H, 2)),       # Normalised [X, Y] in [0, 1]
    'future': np.array(shape=(F, 2)),
    'history_rkt': np.array(shape=(H, 10)),  # 5 keypoints × 2 coords, normalised
    'future_rkt': np.array(shape=(F, 10)),
    'sport': str,
    'match': str,
    'sequence': str,
    'start_frame': int
}
```

**Normalisation**: Ball coordinates are divided by (1920, 1080). Racket keypoints are divided by the same values.

### Split Files (`<sport>/info/train.json`)

JSON list of `[match_id, rally_id]` pairs:

```json
[["match1", "000"], ["match1", "001"], ...]
```

## Generating Data from Scratch

If you have the raw videos, use `DataPreprocess/` scripts in the [code repository](https://github.com/OrcustD/RacketVision/) to prepare all intermediate files:

```bash
cd DataPreprocess

# 1. Extract video frames to JPG
python extract_frames.py --data_root ../data --sport badminton

# 2. Compute median background frame
python create_median.py --data_root ../data --sport badminton

# 3. Generate dataset_info.json and per-sport split files
python generate_dataset_info.py --data_root ../data

# 4. Generate COCO annotations for RacketPose
python generate_coco_annotations.py --data_root ../data
```

## Generating Trajectory Data

After running BallTrack and RacketPose inference, build `data_traj/` PKLs:

```bash
cd TrajPred

# Interpolate short gaps in ball predictions
python linear_interpolate_ball_traj.py --data_root ../data --sport badminton

# Merge racket predictions with ground truth annotations
python merge_gt_with_predictions.py --data_root ../data --sport badminton

# Build PKL dataset
python build_dataset.py --data_root ../data --sport badminton --history 80 --future 20
```

## Citation

If you find this work useful, please consider citing:

```bibtex
@inproceedings{dong2026racket,
  title={Racket Vision: A Multiple Racket Sports Benchmark for Unified Ball and Racket Analysis},
  author={Dong, Linfeng and Yang, Yuchen and Wu, Hao and Wang, Wei and Hou, Yuenan and Zhong, Zhihang and Sun, Xiao},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence (AAAI)},
  year={2026}
}
```