Hugging Face's logo

Datasets:
lpynow
/
MMA-82

Modalities:
Image
Video
Formats:
imagefolder
Size:
10K - 100K
ArXiv:
Libraries:
Datasets
Dataset card Data Studio Files
xet
 Community
2
MMA-82 / README.md
lpynow's picture
lpynow
Update README.md
e90aca2 verified
preview code
|
Raw
History Blame Contribute Delete
14.7 kB
metadata 
gated: true
extra_gated_heading: MMA-82 Dataset Access Request
extra_gated_description: >-
  Please complete all required fields. If a field is not applicable to your
  case, please enter N/A. Access is intended for legitimate academic,
  educational, and non-commercial research purposes only. The authors may
  manually review each request before granting access.
extra_gated_button_content: Submit access request
extra_gated_prompt: >-
  By requesting access to MMA-82, you agree to use the dataset only for
  legitimate, ethical, non-commercial research or education purposes. You agree
  not to use the dataset for harmful applications, subject re-identification,
  surveillance, profiling, discrimination, surveillance systems, commercial
  products, or any experiment that may cause harm to human subjects. You agree
  not to redistribute, sublicense, publish, or share the dataset with any third
  party. You also agree to comply with applicable institutional, ethical, and
  legal requirements. If MMA-82 is useful for your research, please consider
  starring our GitHub repository and liking our Hugging Face dataset page.
extra_gated_fields:
  Full Name: text
  Affiliation: text
  Department or Laboratory: text
  Country: country
  Institutional Email: text
  Position:
    type: select
    options:
      - Faculty or Principal Investigator
      - Postdoctoral Researcher
      - PhD Student
      - Master Student
      - Undergraduate Student
      - Research Scientist
      - Engineer
      - Industry Researcher
      - Other
  Principal Investigator Name: text
  Principal Investigator Email: text
  Research Topic: text
  Research Purpose: text
  Intended Use:
    type: select
    options:
      - Academic Research
      - Education
      - Benchmark Evaluation
      - Method Development
      - Reproducibility Study
      - Non-commercial Research
      - Other
  Dataset Subset Requested:
    type: select
    options:
      - MMA-82-Rec
      - MMA-82-Det
      - Both MMA-82-Rec and MMA-82-Det
      - Not sure yet
  Non-commercial use only: checkbox
  No redistribution or sharing: checkbox
  No re-identification: checkbox
  No surveillance or profiling use: checkbox
  No harmful human-subject experiments: checkbox
  Compliance with ethical and legal requirements: checkbox
  Agreement to cite MMA-82: checkbox

MMA-82: A New Multi-Domain Benchmark for Micro-Action Recognition and Detection
Submitted to IEEE TMM 2026

Yanbin Hao*, Pengyu Liu*, Xing Wei, Xun Yang, Dan Guo, Meng Wang

Hefei University of Technology   |   University of Science and Technology of China
* Equal contribution

📄 Paper   |   📚 arXiv   |   🏠 Project Page   |   🤗 Dataset   |   ⭐ Code

📧 If you have any questions, please feel free to contact me: lpynow@gmail.com

📢 News

  • 🚀2026.07.02: Release of the all dataset.
  • 2026.06.23: Release of the Lab Interview Videos subset of the MMA-82-Rec dataset.
  • 2026.06.20: MMA-82 paper, project page, and dataset links are available.
  • 2026.06.20: We introduce MMA-82, a multi-domain benchmark for micro-action recognition and multi-label temporal detection.

🔎 Overview

MMA-82 extends micro-action analysis from controlled laboratory settings to realistic multi-domain scenarios. It expands the previous MA-52 label space from 52 to 82 fine-grained whole-body micro-action categories and covers four domains: laboratory interviews, psychiatric patient interviews, street interviews, and emotion-rich television videos.

MMA-82 is designed around two core tasks:

  • Micro-Action Recognition: classify trimmed clips into fine-grained body-level and action-level micro-action categories.
  • Multi-label Micro-Action Detection: localize and classify all micro-action instances in untrimmed videos, including dense or overlapping subtle actions.

The benchmark further supports in-domain, cross-domain, zero-shot, and few-shot evaluation protocols, making it a challenging testbed for realistic micro-action understanding.

MMA-82 overview

Representative MMA-82 samples

🧭 Quick Navigation

In this repository, we provide:

📦 Dataset Scale

MMA-82 contains 82 action-level categories organized into seven body-level groups: Body, Head, Upper Limb, Lower Limb, Body-Hand, Head-Hand, and Leg-Hand.

Split Videos / Clips Instances Duration Subjects Description
MMA-82-Rec 39,816 clips 39,816 28.94h 454 Trimmed clips for micro-action recognition
MMA-82-Det 11,180 videos 39,758 46.93h 434 Untrimmed videos for multi-label temporal detection
Total - 79,574 75.87h 454 Four domains and 82 action categories

MMA-82 taxonomy

MMA-82-Rec

MMA-82-Rec contains 39,816 trimmed clips. Each clip is annotated with a body-level group and an action-level micro-action category.

Data Source Train Val Test Total Clips Total Duration Avg. Length Subjects
Laboratory Interviews 15,820 5,636 6,053 27,509 20.48h 2.68s 229
Psychiatric Interviews 4,203 1,398 1,410 7,011 5.72h 2.94s 19
Street Interviews 2,358 792 775 3,925 2.10h 1.92s 26
Emotion Videos 795 247 329 1,371 0.64h 1.67s 180
Total 23,176 8,073 8,567 39,816 28.94h 2.62s 454

MMA-82-Rec statistics

MMA-82-Det

MMA-82-Det contains 11,180 untrimmed videos and 39,758 temporal action instances. Each video contains 3.56 micro-action instances on average, and each instance lasts approximately 3.66s.

MMA-82-Det statistics

🧪 Evaluation Protocols

Micro-Action Recognition

Given a trimmed video clip, the model predicts the target micro-action category at both action and body levels.

  • In-domain setting: train, validation, and test splits come from the same source domain.
  • Cross-domain zero-shot setting: train on laboratory interviews and test directly on another domain.
  • Cross-domain few-shot setting: train on laboratory interviews and adapt with a small number of labeled target-domain samples.
  • Metrics: Top-1 Acc, Top-5 Acc, mean class accuracy (MCA), Macro F1, and Micro F1.

Multi-label Micro-Action Detection

Given an untrimmed video, the model predicts a set of temporal action proposals:

(start_time, end_time, action_category, confidence_score)
  • Goal: detect every micro-action instance and classify its category.
  • Challenge: actions are short, subtle, dense, and may co-occur in rapid succession.
  • Metrics: Detection-mAP at multiple temporal IoU thresholds, reported at both action and body levels.

📊 Baseline Results

Recognition: In-Domain Summary

The paper evaluates skeleton-based PoseC3D and RGB-based GC-TSM baselines. The table below reports action-level test metrics.

Sub-Dataset Method Top-1 Acc Top-5 Acc MCA Macro F1
MMA-82-Rec (All) Skeleton 56.62 80.45 36.77 39.39
MMA-82-Rec (All) RGB 60.43 86.14 38.98 39.56
Laboratory Interviews Skeleton 64.84 87.81 44.88 47.03
Laboratory Interviews RGB 68.15 92.83 48.01 46.48
Psychiatric Interviews Skeleton 41.63 68.79 15.74 17.65
Psychiatric Interviews RGB 46.74 72.13 14.10 13.64
Street Interviews Skeleton 40.77 70.71 19.49 20.70
Street Interviews RGB 39.87 73.55 12.75 12.56
Emotion Videos Skeleton 7.29 17.93 3.72 3.11
Emotion Videos RGB 25.53 52.89 12.20 11.05

Recognition: Cross-Domain Summary

PoseC3D is trained on laboratory interviews and evaluated on target domains under zero-shot and few-shot protocols.

Target Domain Protocol Action Top-1 Action Top-5 Action MCA Body Top-1
Psychiatric Interviews Zero-Shot 27.30 50.28 14.25 68.58
Psychiatric Interviews 1-Shot 30.27 58.62 22.82 72.48
Psychiatric Interviews 5-Shot 30.12 58.60 22.72 72.60
Psychiatric Interviews 10-Shot 30.13 58.59 22.71 72.64
Street Interviews Zero-Shot 20.65 44.90 10.65 53.16
Street Interviews 1-Shot 21.38 45.64 15.60 54.95
Street Interviews 5-Shot 21.58 46.10 16.17 64.91
Street Interviews 10-Shot 22.52 45.92 17.76 65.64
Emotion Videos Zero-Shot 14.13 32.98 6.63 43.92
Emotion Videos 1-Shot 17.26 39.77 14.00 41.43
Emotion Videos 5-Shot 17.48 41.75 15.20 42.35
Emotion Videos 10-Shot 17.70 42.53 15.43 42.03

Detection: AdaTAD on MMA-82-Det

Backbone Action mAP@0.2 Action mAP@0.5 Action mAP@0.7 Action Avg Body mAP@0.2 Body mAP@0.5 Body mAP@0.7 Body Avg AVG
VideoMAE-S 20.88 12.72 5.56 12.09 48.18 28.78 13.91 25.44 18.77
VideoMAE-B 22.62 14.67 6.32 13.59 50.95 30.46 12.23 29.13 21.36
VideoMAE-L 22.74 15.60 7.68 14.98 55.48 33.01 14.06 31.83 23.41
VideoMAE-H 26.53 17.56 7.55 16.08 54.71 33.64 14.17 30.05 23.07

MMA-82 recognition examples

MMA-82 detection examples

🎭 Emotion and Micro-Actions

MMA-82 also studies how micro-actions relate to affective states. The paper shows that micro-actions are strongly associated with emotions and provide complementary cues beyond facial micro-expressions.

Setting Method Top-1 Acc F1
Micro-Expression Only DeepFace 22.86 17.54
Micro-Action Only TSM 32.38 31.86
Both DeepFace + TSM 32.86 32.36

Key observations:

  • Sad and melancholy both correlate with bowing head and turning head.
  • Sad contains more explicit negative bodily actions, while melancholy is subtler and more inward.
  • Micro-actions alone outperform facial micro-expression cues under the reported setup.
  • Combining micro-actions and micro-expressions improves over the facial baseline.

Micro-actions and emotion Sankey diagram

Emotion-rich MMA-82 examples

📥 Dataset

The dataset page is available on Hugging Face:

hf download lpynow/MAR_plus_plus \
  --repo-type dataset \
  --local-dir data/MMA-82

Due to redistribution restrictions, part of the Emotion, Psychiatric Interview, and Street Interview videos are released as rebuild metadata instead of raw videos. After downloading the required original sources, users can reconstruct the restricted-source subsets with the unified release script:

pip install -r MMA82_Release/requirements.txt

python MMA82_Release/mma82_release.py prepare \
  --task both \
  --domain all \
  --jobs 4 \
  --output-root data/MMA-82

For users who have already prepared the original videos locally:

python MMA82_Release/mma82_release.py prepare \
  --caer-dir /path/to/CAER \
  --youtube-source-root /path/to/youtube_sources \
  --skip-download \
  --task both \
  --domain all \
  --jobs 4 \
  --output-root data/MMA-82

youtube_sources should contain files named P001.mp4, V001.mp4, etc. Some YouTube videos may require cookies or age/login confirmation; the script supports passing yt-dlp options such as --yt-dlp-cookies and --yt-dlp-cookies-from-browser.

Please see MMA82_Release/README_REBUILD.md for the full rebuild instructions, requirements, source-file layout, and restricted/direct-release split of MMA-82-Det.

📑 Citation

If you find MMA-82 useful for your research, please consider citing our paper:

@misc{hao2026newmultidomainbenchmarkmicroaction,
  title={A New Multi-Domain Benchmark for Micro-Action Recognition and Detection},
  author={Hao, Yanbin and Liu, Pengyu and Wei, Xing and Yang, Xun and Guo, Dan and Wang, Meng},
  year={2026},
  eprint={2606.14096},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2606.14096}
}

🙏 Acknowledgement

The MMA-82 project page is adapted from the Academic Project Page Template and the Nerfies project page. We also thank the authors of MA-52, MMA-52, PoseC3D, GC-TSM, AdaTAD, VideoMAE, DeepFace, and related micro-action / micro-expression benchmarks for their inspiring work.