Datasets:

juyil
/

AVQA-videos

license: other
pretty_name: AVQA (Audio-Visual QA) — Videos + Annotations
task_categories:
  - visual-question-answering
  - multiple-choice
language:
  - en
tags:
  - audio-visual
  - video-question-answering
  - avqa
  - vggsound
size_categories:
  - 10K<n<100K

AVQA — Audio-Visual Question Answering (videos + annotations)

A drop-in package of the AVQA dataset (Yang et al., ACM MM 2022): real-life audio-visual question answering over short in-the-wild clips. The original release ships only the QA annotations and expects users to collect the source videos from VGGSound themselves. This repository bundles the source video clips together with the official train/val annotations, so the dataset is usable without any YouTube scraping.

What's inside

videos/<video_name>.mp4   # 56,666 clips, ~10 s each, H.264 video + AAC audio
train_qa.json             # 40,425 QA pairs
val_qa.json               # 16,910 QA pairs

Videos are VGGSound clips, named <youtube_id>_<start_seconds> — identical to the video_name field in the annotations, so each question maps directly to a file.
Coverage: 56,666 of the 57,015 referenced clips are present (99.4 % of all QA pairs). The missing ~0.6 % are source clips that are no longer retrievable.

Annotation schema

Each record in train_qa.json / val_qa.json:

field	meaning
`video_name`	clip id; the file is `videos/<video_name>.mp4`
`question_text`	the question
`multi_choice`	list of 4 answer options
`answer`	index (0–3) of the correct option within `multi_choice`
`question_relation`	modality needed to answer: `Sound` / `View` / `Both`
`question_type`	question category: `Happening` / `Come From` / `Which` / `Where` / `Why` / `When` / `Before Next` / `Used For`
`id`, `video_id`	numeric ids from the original release

Usage

import json, os

qa = json.load(open("train_qa.json"))
ex = qa[0]
video   = os.path.join("videos", ex["video_name"] + ".mp4")
options = ex["multi_choice"]
gold    = options[ex["answer"]]
print(ex["question_text"], options, "->", gold, "| modality:", ex["question_relation"])

Download the whole dataset:

hf download juyil/AVQA-videos --repo-type dataset --local-dir AVQA

Provenance & credits

Annotations are from the original AVQA release (GitHub: AlyssaYoung/AVQA). All annotation rights belong to the original authors.
Videos are clips from VGGSound (Chen et al., ICASSP 2020), which are segments of public YouTube videos. Copyright of the underlying footage remains with the original uploaders.

Citation

@inproceedings{yang2022avqa,
  title     = {AVQA: A Dataset for Audio-Visual Question Answering on Videos},
  author    = {Yang, Pinci and Wang, Xin and Duan, Xuguang and Chen, Hong and
               Hou, Runze and Jin, Cong and Zhu, Wenwu},
  booktitle = {Proceedings of the 30th ACM International Conference on Multimedia},
  year      = {2022}
}

@inproceedings{chen2020vggsound,
  title     = {VGGSound: A Large-scale Audio-Visual Dataset},
  author    = {Chen, Honglie and Xie, Weidi and Vedaldi, Andrea and Zisserman, Andrew},
  booktitle = {ICASSP},
  year      = {2020}
}

Disclaimer

Provided for non-commercial research only. The video segments originate from YouTube via VGGSound; all rights remain with their respective owners. If you are a rights holder and would like content removed, please open a discussion on this repository.