README.md

---
license: cc-by-nc-sa-4.0
language:
- en
pretty_name: AVSCapBench
tags:
- video-captioning
- audio-visual
- omni-modal
- event-recall
- benchmark
size_categories:
- 1K<n<10K
---

# AVSCapBench

[![Project Page](https://img.shields.io/badge/Project%20Page-AVSCap-1B2838?logo=githubpages&logoColor=white)](https://nju-link.github.io/AVSCap/)
&nbsp;
[![Paper](https://img.shields.io/badge/arXiv-2606.xxxxx-b31b1b)](https://arxiv.org/abs/2606.xxxxx)
&nbsp;
[![GitHub](https://img.shields.io/badge/GitHub-AVSCap-181717?logo=github&logoColor=white)](https://github.com/NJU-LINK/AVSCap)

AVSCapBench contains 1,226 manually annotated omni-modal video clips. Each sample includes a dense caption, visual events, audio events split into `speech`, `music`, and `sfx`, and audio-visual synergistic events.

## Download

```bash
hf download NJU-LINK/AVSCapBench --repo-type dataset --local-dir AVSCapBench
```

## Structure

```text
videos/
  1.mp4
  2.mp4
  ...
metadata.jsonl
metadata/
  OmniCaption.json
```

`metadata.jsonl` is provided for the Hugging Face Dataset Viewer. It links each video with its annotations through the `file_name` field.

## AVSCapBench

AVSCapBench contains 1,226 manually annotated omni-modal video clips. The dataset release includes the videos and the complete ground-truth file `OmniCaption.json`.

Each annotation item has the following structure:

```json
{
  "video_id": 1,
  "video_path": "./videos/1.mp4",
  "caption": "...",
  "event": {
    "visual_events": [],
    "audio_events": {
      "speech": [],
      "music": [],
      "sfx": []
    },
    "synergistic_events": []
  },
  "visual_caption": "..."
}
```

<p align="center">
  <img src="assets/avscapbench_statistics.png" width="92%" alt="AVSCapBench statistics">
</p>

AVSCapBench provides longer, denser, and more audio-rich captions than existing audio-visual captioning benchmarks.

<p align="center">
  <img src="assets/benchmark_comparison.png" width="92%" alt="Benchmark comparison">
</p>

## Main Annotation File

`metadata/OmniCaption.json` is the full ground-truth file. Each record contains:

- `video_id`: numeric id matching `videos/{video_id}.mp4`
- `video_path`: relative video path
- `caption`: dense omni-modal caption
- `event.visual_events`: visual checklist
- `event.audio_events.speech`: speech checklist
- `event.audio_events.music`: music checklist
- `event.audio_events.sfx`: sound-effect checklist
- `event.synergistic_events`: audio-visual binding checklist
- `visual_caption`: visual-only caption

## Use With Evaluation Code

Use this dataset folder together with the GitHub release folder:

```bash
python evaluation/evaluate_avscapbench.py \
  --gt AVSCapBench/metadata/OmniCaption.json \
  --videos-dir AVSCapBench/videos \
  --captions-dir model_captions \
  --output-dir results/eval
```

## License

Our dataset is under the CC-BY-NC-SA-4.0 license.