FürElise: Capturing and Physically Synthesizing Hand Motions of Piano Performance

This hosts the FürElise dataset, which contains 10 hours, 153 pieces of piano music performed by 15 elite-level pianists, along with synchronized audio and key pressing events.

Getting Started

To download the dataset and the related scripts, first run

git lfs install
# We skip the raw dataset.zip because it's too large
GIT_LFS_SKIP_SMUDGE=1 git clone  git@hf.co:datasets/rcwang/for_elise

And then download assets, which will download around 44G of data.

cd for_elise
sh ./download_data.sh

3D Visualizer

We provide a web-based 3D visualizer of our dataset under visualizer/ which has the same functionality as the one you see in the project website. You can run it with the following commands:

cd visualizer
python server.py

Then you can access the visualizer by visiting http://127.0.0.1:8080. You need flask and numpy to run the visualizer.

For some long pieces, it might take several seconds to initialize the visualizer.

Dataset structure

The metadata for all 153 pieces can be found in metadata.json:

[
  {
    "name": "Concerto in D Minor, BWV 1052: I. Allegro",
    "composer": "J.S. Bach",
    "piece_id": 0,
    "subject_id": 0
  },
  ...
]

Motion data for each music piece is stored in a single subdirectory under dataset/. The motion data is stored frame by frame with FPS 59.94.

piece_id # Directory
├── motion.pkl                    # Pickle file of a dictionary storing 3D hand motion data
│   ├── left                      # Motion data for the left hand
│   │   ├── joints                # Nx21x3, joint locations for every frame
│   │   ├── mano_params           # MANO hand parameters for each frame
│   │   │   ├── global_translation  # Nx3
│   │   │   ├── global_rotation     # Nx3x3
│   │   │   ├── pose                # Nx45
│   │   │   ├── shape               # 10
│   │   │   └── vertices            # Nx778x3
│   ├── right                     # Motion data for the right hand
│   │   ├── joints                # Nx21x3, joint locations for every frame
│   │   ├── mano_params           # MANO hand parameters for each frame
│   │   │   ├── global_translation  # Nx3
│   │   │   ├── global_rotation     # Nx3x3
│   │   │   ├── pose                # Nx45
│   │   │   ├── shape               # 10
│   │   │   └── vertices            # Nx778x3
├── midi.mid                      # Synchronized MIDI file recorded during performance
├── audio.mp3                     # Synchronized audio synthesized from the MIDI file
└── vis                           # Used for the 3D visualizer
    ├── metadata.json
    └── pressed_keys.pkl

Licensing

This dataset is released under the CC BY-NC 4.0 license.

Cite

@inproceedings{wang2024piano,
  title = {FürElise: Capturing and Physically Synthesizing Hand Motions of Piano Performance},
  author = {Ruocheng Wang and Pei Xu and Haochen Shi and Elizabeth Schumann and C. Karen Liu},
  booktitle = {SIGGRAPH Asia 2024},
  year = {2024}
}