gouba2333/BoxComm-Dataset

# BoxComm-Dataset

BoxComm-Dataset is the official data release for BoxComm, a benchmark for category-aware boxing commentary generation and narration-rhythm evaluation.

## Resources

- Project Page: https://gouba2333.github.io/BoxComm
- Paper: http://arxiv.org/abs/2604.04419
- Code: https://github.com/gouba2333/BoxComm
- Benchmark: https://huggingface.co/datasets/gouba2333/BoxComm

## Overview

This dataset release is intended for training, analysis, and reproducible preprocessing. It contains the complete processed videos together with the released annotations and benchmark metadata.

Recommended structure:

```text
BoxComm-Dataset/
├── train/
│ ├── videos/
│ ├── events/
│ └── asr/
├── eval/
│ ├── videos/
│ ├── events/
│ └── asr/
└── metadata/
```

The split convention is:

- `train`: video id `< 478`
- `eval`: video id `>= 478`

Each event directory should contain:

- one skeleton `.pkl` file
- one `video_event_inference_3.json` file

Each ASR JSON file should contain `classified_segments`.

## What is included

- processed match videos
- event annotations
- skeleton data
- ASR with sentence segmentation
- 3-way commentary labels
- split metadata

## Intended uses

- supervised fine-tuning for commentary generation
- category-aware commentary evaluation
- narration-rhythm analysis
- multimodal sports video understanding research

## Data preparation in the code repository

The official code repository provides:

- `scripts/prep_qwen3vl_sft_data.py`
- `scripts/train_qwen3vl.py`
- `scripts/infer_qwen3vl.py`
- `scripts/eval_metrics.py`
- `scripts/eval_streaming_cls_metrics.py`

Repository: https://github.com/gouba2333/BoxComm

## Licensing

The public release includes processed videos, ASR annotations, event JSON files, skeleton PKL files, and benchmark metadata for research use.

## Citation

```bibtex
@article{wang2026boxcomm,
title={BoxComm: Benchmarking Category-Aware Commentary Generation and Narration Rhythm in Boxing},
author={Wang, Kaiwen and Zheng, Kaili and Deng, Rongrong and Shi, Yiming and Guo, Chenyi and Wu, Ji},
journal={arXiv preprint arXiv:2604.04419},
year={2026}
}
```

# BoxComm数据集（BoxComm-Dataset）

BoxComm数据集是BoxComm基准测试的官方数据发布版本，该基准面向类别感知拳击解说生成与解说节奏评估任务。

## 资源
- 项目页面：https://gouba2333.github.io/BoxComm
- 论文：http://arxiv.org/abs/2604.04419
- 代码：https://github.com/gouba2333/BoxComm
- 基准数据集：https://huggingface.co/datasets/gouba2333/BoxComm

## 数据集概览
本数据集发布版本旨在支持模型训练、分析与可复现预处理工作，包含完整的处理后赛事视频、已发布标注以及基准元数据。

推荐目录结构如下：
text
BoxComm-Dataset/
├── train/
│ ├── videos/
│ ├── events/
│ └── asr/
├── eval/
│ ├── videos/
│ ├── events/
│ └── asr/
└── metadata/

数据集划分规则如下：
- `train`：视频ID `< 478`
- `eval`：视频ID `>= 478`

每个事件目录需包含：
- 一个骨架数据.pkl文件
- 一个`video_event_inference_3.json`文件

每个自动语音识别（Automatic Speech Recognition，ASR）JSON文件应包含`classified_segments`字段。

## 数据集包含内容
- 处理后的赛事视频
- 事件标注
- 骨架数据
- 带分句的ASR结果
- 三分度解说标签
- 划分元数据

## 适用场景
- 解说生成任务的监督微调
- 类别感知解说评估
- 解说节奏分析
- 多模态体育视频理解研究

## 代码仓库中的数据预处理脚本
官方代码仓库提供以下脚本：
- `scripts/prep_qwen3vl_sft_data.py`
- `scripts/train_qwen3vl.py`
- `scripts/infer_qwen3vl.py`
- `scripts/eval_metrics.py`
- `scripts/eval_streaming_cls_metrics.py`

仓库地址：https://github.com/gouba2333/BoxComm

## 许可协议
本公开发布版本包含处理后视频、ASR标注、事件JSON文件、骨架PKL文件以及基准元数据，仅可用于科研用途。

## 引用格式
bibtex
@article{wang2026boxcomm,
title={BoxComm: Benchmarking Category-Aware Commentary Generation and Narration Rhythm in Boxing},
author={Wang, Kaiwen and Zheng, Kaili and Deng, Rongrong and Shi, Yiming and Guo, Chenyi and Wu, Ji},
journal={arXiv preprint arXiv:2604.04419},
year={2026}
}