junhob/v2a-onset-audit

V2A Onset Alignment Audit

Data for Training Refines Onset Placement but Erodes Event Correspondence: A Full-Trajectory Audit of Video-to-Audio Onset Alignment — WACV 2027.

JoonHo Byun · Taegon Kim · Juyeon Park · Unsang Park, Sogang University.

Code and evaluation instructions

Contents

Component	Included
V266 benchmark	266 clip IDs, 2,225 full-clip onset annotations, class labels, source IDs/start times, and detector-rule split
Greatest Hits	Evaluation split IDs and 244 onset annotation files with original attribution
Curation	178,419 training-clip scores/status records and 140,012 retained IDs
Measurements	99 checkpoint conditions, validation scores, envelope instruments, and separate legacy trajectories
Analysis features	Detected onsets and spectral-flux envelopes for all 99 conditions × 510 clips
Representative generated audio	Vanilla 10K and 150K, each with 244 GH and 266 V266 clips: 1,020 FLAC files total

This compact release contains no source videos, original recorded audio, or model weights. The FLAC files are model-generated outputs, not recordings extracted from the source videos. Generated waveforms for the other conditions are not part of this release; their predictions, envelopes, and archived measurements remain available.

Source media: links and identifiers

• Greatest Hits: obtain original media from the authors' Visually Indicated Sounds project. metadata/gh_test_ids.txt and metadata/gh_validation_ids.txt identify the evaluation subsets.
• VGGSound: this release provides file stems (clip_id), YouTube IDs, and segment start times in metadata/v266_clip_metadata.csv, plus our onset annotations. Original media are not mirrored. See the VGGSound project and author-maintained repository.

Clip identifiers and annotations do not grant rights to the underlying media. Follow the original providers' access conditions and licenses when obtaining recordings. Source videos may become unavailable; identifiers are retained without replacing missing videos with other clips.

Download and use

hf download junhob/v2a-onset-audit --repo-type dataset --local-dir v2a-onset-data

From the code repository:

python scripts/setup_data.py --release-root /path/to/v2a-onset-data --data-root data \
  --audio-shard audio_vanilla_10000.tar --audio-shard audio_vanilla_150000.tar

The helper verifies hashes before extraction. Omit the audio options to install only benchmark metadata and GH annotations. See metadata, metrics, and assets for schemas and protocol details. release_manifest.json lists the files and SHA-256 checksums; its own contents are excluded from that list.

Scope and limitations

V266 samples an audio-visually aligned subset across 100 classes. Its target is an audible onset with an identifiable visible cause. Event scoring uses the first eight seconds: 243 GH and 263 V266 clips have nonempty reference windows. Empty-reference windows are excluded, so these scores do not measure false positives on globally onset-free clips.

The 99 conditions are checkpoints and inference controls, not independent training seeds. The newer EMA grid and older full-budget measurements have distinct protocols and namespaces. The earlier submitted 90/85 snapshot has no recovered exact membership list; this release retains the available 99 records with explicit scope flags.

Features were recomputed from retained generated outputs during release preparation. Their event-score aggregates match 594/594 archived values at four decimals. This does not reproduce model training, waveform generation, or distributional-quality feature extraction. Onset scores and envelope measurements are not direct tests of perceptual quality or absence.

License and citation

See component licenses. Cite this paper and the original datasets as applicable.

@inproceedings{jhbyun2027syncfreeze,
  title={Training Refines Onset Placement but Erodes Event Correspondence:
         A Full-Trajectory Audit of Video-to-Audio Onset Alignment},
  author={Byun, JoonHo and Kim, Taegon and Park, Juyeon and Park, Unsang},
  booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  year={2027}
}