Mimics & Muscles — Datasets
Participants performed standardized facial movements and emotional expressions while surface electromyography and a frontal camera recorded together. Every clip is cut to one task, and the sEMG is resampled to the video frame rate, so signal and image line up by row index. MMM and MaM follow the identical recording and processing pipeline, and differ only in scale and access.
01 Two companion datasets
Same recording setup, same movements, same processing pipeline — released at two different scales and access levels.
MMM — Mimics, Muscles, and Me
1 participant · sEMG + frontal video · free download, no request needed.MaM — Mimics and Muscles
20 participants · same format and movements as MMM · available on request.02 What is happening here?
Each participant sat in front of a camera and did two kinds of things, twice each — once with sEMG electrodes on the face, once without. The electrode array measures muscle activity but corrupts the face for any image-based method, so the pipeline does three jobs on top of the raw recording:
sEMG + frontal video
11 facial muscles × 2 sides at 4096 Hz, and a 286×286 / 30 fps camera, running on independent clocks.Frame-align the modalities
Consensus cross-correlation of sEMG envelopes vs. video blendshapes finds the lag; clips are cut so sEMG row i = video frame i.Remove electrodes → 3DMM
MC-CycleGAN synthesizes electrode-free video; seven monocular 3D face reconstructors are fitted to it.
With sEMG electrodes — raw frontal video
Electrodes removed — MC-CycleGAN03 Two modalities, perfectly aligned
There was no hardware trigger LED, so alignment is statistical. After the cut, the sEMG envelope (22 channels, 11 muscles × L/R) shares one time axis with the video — the dashed marker in the heatmap below is the exact frame shown beside it. The example below is drawn from the public MMM release; MaM follows the same alignment procedure.
video frame 105
99_2S_emotion_12_happy
22-channel sEMG envelope · the marker = that frame04 What’s in the box
MMM and MaM share the same directory layout and file schema — only the number of participants and clips differs. Each release ships a CSV index that is the authoritative inventory; iterate over that first.
| Directory | Contents |
|---|---|
emg/ |
Per-clip sEMG CSV, one row per video frame, 22 channels |
video/ |
Frontal video, cut to the synced window |
video_no-cut/ |
Same clips before the sync cut (lead-in/out intact) |
video_noelec/ |
Electrode-removed version of video/ (MC-CycleGAN) |
video_noelec_params/ |
3DMM params fitted to video_noelec/, one folder per method |
*.csv |
Per-release clip index (authoritative inventory) |
Seven 3D morphable-model methods
Per-frame parameters for every clip with sEMG; each file has one row per video frame, aligned to video frame i and sEMG row i.
05 Citation & license
Use requires citing the three papers below. The MC-CycleGAN paper covers the electrode-removed videos that every 3DMM parameter set is fitted to. Processing code lives in the SensorGAN repo.
EIFER — CVPR 2025
@inproceedings{buechner2025electromyography,
doi = {10.1109/CVPR52734.2025.00029},
year = {2025},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
title = {Electromyography-Informed Facial Expression Reconstruction for Physiological-Based Synthesis and Analysis},
author = {Tim Büchner and Christoph Anders and Orlando Guntinas-Lichius and Joachim Denzler},
}
MC-CycleGAN — ACIVS 2023
@inproceedings{buechner2023improved,
doi = {10.1007/978-3-031-45382-3_22},
pages = {262-274},
year = {2023},
booktitle = {Advanced Concepts for Intelligent Vision Systems (Acivs)},
author = {Tim Büchner and Orlando Guntinas-Lichius and Joachim Denzler},
title = {Improved Obstructed Facial Feature Reconstruction for Emotion Recognition with Minimal Change CycleGANs},
}
sEMG recording protocol — Scientific Reports 2023
@article{guntinas2023high,
title={High-resolution surface electromyographic activities of facial muscles during the six basic emotional expressions in healthy adults: a prospective observational study},
author={Guntinas-Lichius, Orlando and Trentzsch, Vanessa and Mueller, Nadiya and Heinrich, Martin and Kuttenreich, Anna-Maria and Dobel, Christian and Volk, Gerd Fabian and Gra{\ss}me, Roland and Anders, Christoph},
journal={Scientific reports},
volume={13},
number={1},
pages={19214},
year={2023},
publisher={Nature Publishing Group UK London}
}