Dataset Card for VidMuse-V2M

This is a FiftyOne grouped dataset with 300 groups (600 samples).

This FiftyOne dataset contains only the 300 video-music pairs of the V2M-bench subset, which is the only part of V2M whose media is distributed in the source repo.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("harpreetsahota/VidMuse-V2M")

# Launch the App
session = fo.launch_app(dataset)

Dataset Details

Dataset Description

V2M is a video-to-music dataset proposed in the VidMuse project, built from YouTube videos of genres such as movie trailers, advertisements, documentaries and vlogs. The paper describes three subsets: V2M (360K video-music pairs, ~1.8x10^4 hours, for pretraining), V2M-20K (20K pairs, ~6x10^2 hours, for finetuning) and V2M-bench (300 pairs, 9 hours, for evaluation). The source Hugging Face repo ships media only for V2M-bench; V2M and V2M-20K are provided as lists of YouTube IDs. This FiftyOne dataset is built from the V2M-bench media.

• Curated by: Zeyue Tian, Zhaoyang Liu, Ruibin Yuan, Jiahao Pan, Qifeng Liu, Xu Tan, Qifeng Chen, Wei Xue, Yike Guo (HKUST Audio)
• Funded by: [More Information Needed]
• Shared by: [More Information Needed]
• Language(s): en
• License: CC BY-NC 4.0

Dataset Sources

• Repository: https://huggingface.co/datasets/HKUSTAudio/VidMuse-V2M-Dataset
• Paper: https://arxiv.org/abs/2406.04321
• Demo: [More Information Needed]

Uses

Direct Use

Research on video-to-music generation, including benchmarking models on V2M-bench.

Out-of-Scope Use

The data is licensed CC BY-NC 4.0, so commercial use is not permitted. [More Information Needed] for other out-of-scope uses.

Dataset Structure

This is a FiftyOne grouped dataset (media_type="group") with 300 groups, one per YouTube video, and 600 samples. Each group has two slices:

Slice	Media type	Description
video (default)	video	Silent .mp4 video, h264, 720 px tall (297 are 1280x720; the other three are 1194x720, 1274x720 and 1278x720), 24-30 fps. The files have no audio stream.
music	unknown	The paired music track as a .wav file (pcm_s16le, 44.1 kHz, stereo)

FiftyOne has no native audio media type, so the .wav files are stored as unknown samples in their own group slice. The App does not play or render these samples by default. The total duration is about 9.01 hours for the videos and about 9.01 hours for the music (measured with ffprobe).

Fields (present on samples in both slices):

Field	FiftyOne type	Description
filepath	StringField	Path to the .mp4 (video slice) or .wav (music slice)
group	Group	Group membership and slice name (video or music)
tags	ListField(StringField)	bench for all samples
metadata	VideoMetadata / Metadata	Duration, frame rate and resolution (video) or file metadata (music), computed with compute_metadata()
youtube_id	StringField	YouTube video ID, which is the filename stem of both files
subset	StringField	V2M subset name; always V2M-bench here

There are no per-sample labels in the source release (no captions or genre labels), so the dataset has no label fields.

dataset.info: source repo URL, paper URL, license, subset_id_counts (V2M: 359,319 IDs, V2M-20k: 20,000 IDs, V2M-bench: 300 IDs) and a note that media is only available for V2M-bench.

Parsing decisions:

• Video and music files are paired by filename stem. All 300 stems match in both directories and match the IDs in V2M-bench.txt exactly.
• The V2M-20k and V2M lists are ID lists with no media, so only their counts are recorded in dataset.info. The V2M-bench IDs are disjoint from both lists.
• The video is the default slice because the videos have no audio, so the music slice is the only source of audio.

Dataset Creation

Curation Rationale

Existing video-music pair datasets were limited in scale, quality or genre diversity, so the authors built a large-scale dataset together with a benchmark for video-to-music generation.

Source Data

Data Collection and Processing

Video-music pairs were collected from YouTube, including data from existing resources such as HIMV-200K and YouTube-8M subsets tagged "Music" and "Trailer". The authors then applied a multi-step pipeline:

1. Coarse filtering: removed videos missing audio or video tracks, videos that were too short or too long, inappropriate content such as violence or explicit material, and categories such as Interview, News and Gaming.
2. Fine-grained filtering: kept videos with substantial music content (PANNs sound event detection; at least 50% of audio frames classified as music at a 0.5 confidence threshold) and removed videos made up mainly of static images (SSIM threshold of 0.8).
3. Music source separation: isolated the music component by removing speech and singing.
4. Audio-visual alignment ranking: ranked videos by audio-visual alignment score to keep strongly correlated pairs, then split the data into V2M, V2M-20K and V2M-bench subsets.

Human experts curated the benchmark subset.

Who are the source data producers?

The creators and uploaders of the YouTube videos. [More Information Needed] on any further demographic details.

Annotations

Annotation process

The source release contains no per-sample annotations. The V2M-bench subset was curated by human experts from the filtered data; the paper does not describe further annotation.

Who are the annotators?

Human experts curated the benchmark subset. [More Information Needed] for further details.

Personal and Sensitive Information

[More Information Needed]

Citation

BibTeX:

@article{tian2024vidmuse,
  title={Vidmuse: A simple video-to-music generation framework with long-short-term modeling},
  author={Tian, Zeyue and Liu, Zhaoyang and Yuan, Ruibin and Pan, Jiahao and Liu, Qifeng and Tan, Xu and Chen, Qifeng and Xue, Wei and Guo, Yike},
  journal={arXiv preprint arXiv:2406.04321},
  year={2024}
}

APA:

Tian, Z., Liu, Z., Yuan, R., Pan, J., Liu, Q., Tan, X., Chen, Q., Xue, W., & Guo, Y. (2024). VidMuse: A simple video-to-music generation framework with long-short-term modeling. arXiv preprint arXiv:2406.04321.

More Information

The source repo lists three ID files: V2M.txt, V2M-20k.txt and V2M-bench.txt.

Dataset Card Authors

[More Information Needed]

Dataset Card Contact

[More Information Needed]