emg2pose Dataset Overview

Dataset Description

The emg2pose dataset contains paired surface electromyography (sEMG) recordings and hand motion‑capture recordings. It comprises 25,253 HDF5 files, each providing time‑aligned 2 kHz sEMG data and joint‑angle data of a single hand in a single stage. Each stage lasts about 1 minute, with 193 participants covering 370 hours across 29 stages.

Dataset Statistics

The statistical information of the dataset is as follows:

Dataset statistics

Metadata

The metadata.csv file includes the following information for each HDF5 file:

Column Name	Description
`user`	Anonymous user ID
`session`	Recording session (each session contains multiple stages)
`stage`	Stage name
`side`	Hand side (`left` or `right`)
`moving_hand`	Whether the hand is prompted to move in this stage
`held_out_user`	Whether the user is excluded from the training set
`held_out_stage`	Whether the stage is excluded from the training set
`split`	Dataset split (`train`, `test`, or `val`)
`generalization`	Generalization type (cross‑user `user`, cross‑stage `stage`, or cross‑user‑and‑stage `user_stage`)

Dataset Versions

Full Dataset: 431 GiB, containing 25,253 HDF5 files.
Small Dataset: Approximately 600 MiB for quick start‑up and validation.

Pre‑trained Models

Checkpoint files are provided for the following pre‑trained models:

vemg2pose (tracking and regression setup)
neuropose (regression setup)

License

The emg2pose dataset is released under the CC‑BY‑NC‑SA‑4.0 license.

emg2pose

Description

emg2pose Dataset Overview

Dataset Description

Dataset Statistics

Metadata

Dataset Versions

Pre‑trained Models

License

AI studio

Access Dataset

Topics

Source