Datasets:
license: cc-by-nc-4.0
pretty_name: PianoVid-R
task_categories:
- video-classification
tags:
- piano
- music-transcription
- fingering
- audio-visual
size_categories:
- 10K<n<100K
extra_gated_prompt: >-
PianoVid-R is released under CC BY-NC 4.0 for non-commercial research. By
requesting access you agree not to attempt to identify the pianists, not to
redistribute the videos, and to cite the paper in work that uses the dataset.
extra_gated_fields:
Name: text
Affiliation: text
I agree to the terms above: checkbox
PianoVid-R
Video recordings of piano performances from 80 pianists, paired with the MIDI recorded by the piano and fingering labels derived from the video. The dataset accompanies the paper A Video Dataset and Key-Aligned Model for Visual Piano Transcription and Fingering Recognition (Liu, Chen, Grüner, Koepke).
Paper: https://arxiv.org/abs/XXXX.XXXXX · Project page: https://timbermorph.github.io/PianoLite/ · Code: https://github.com/Timbermorph/PianoLite
All recordings were made on one Yamaha Disklavier under a fixed overhead camera, so onsets, key releases and velocities come from the instrument rather than from audio transcription. Video and MIDI are aligned to within one frame for almost every recording. Fingering labels are derived from the video by reconstructing the hands in 3D and matching fingertips to the keys at each onset; a manual check agrees with 95.6% of them. All pianists consented to the recording and its release.
| Hours | Pianists | Onsets | Note labels | Finger labels | |
|---|---|---|---|---|---|
| PianoVid-R | 39.6 | 80 | 814,554 | recorded MIDI | 98% of onsets |
| Split | Recordings | Pianists | Hours | Onsets | Onsets / s |
|---|---|---|---|---|---|
| Train | 80 | 58 | 25.8 | 398k | 4.3 |
| Val | 12 | 10 | 5.6 | 145k | 7.2 |
| Test | 18 | 12 | 8.2 | 272k | 9.2 |
Layout
Rectified 800 x 144 grayscale video at 25 fps in 15-second chunks (375 frames, consecutive chunks overlap by 4 frames), one per-frame label file per chunk, one note list per recording, and one fingering file per chunk.
dataset.csv one row per chunk
videos/<recording>/chunk_<i>.mp4
annotations/<recording>/chunk_<i>.npz
events/<recording>.tsv
finger_labels/<recording>/chunk_<i>.npz
dataset.csv columns: split (train / val / test), chunk_path, annotation_path,
record_time (recording id), start_frame (first frame of the chunk in the recording),
num_frames, fps, event_path, label_start_frame (the chunk's first frame on the
timeline of the event file).
Annotation files hold [num_frames, 88] arrays: onset_roll (1 at the onset frame),
frame_roll (1 while the key is held), offset_roll (1 at the key-release frame),
velocity_roll (MIDI velocity at the onset frame) and velocity_mask. Event files are
tab-separated with the header # onset key_offset frame_offset note velocity (seconds;
key_offset is the physical key release). Fingering files hold finger_roll
[num_frames, 88] with 0 = no onset, 1-5 = left thumb to little finger, 6-10 = right thumb
to little finger, 255 = unknown.
Use
Download into dataset/pianovid_r/ of the PianoLite
repository and set dataset.data_root in a config to that directory.
License
CC BY-NC 4.0, for non-commercial research. Please do not attempt to identify the pianists, do not redistribute the videos, and cite the paper in work that uses the dataset. If a pianist withdraws consent, the affected recordings are removed in the next release; please delete them from your copy when notified. Contact: jianzhe.liu@tum.de
Citation
@article{liu2026pianolite,
title = {A Video Dataset and Key-Aligned Model for Visual Piano Transcription and Fingering Recognition},
author = {Liu, Jianzhe and Chen, Danqing and Gr{\"u}ner, Inken and Koepke, A. Sophia},
journal = {arXiv preprint arXiv:XXXX.XXXXX},
year = {2026}
}