Thank you for your interest in the MIAM (Multimodal Industrial Activity Monitoring) dataset. To ensure responsible use and track dataset usage, please provide the following information:

By requesting access, you agree to:

Use this dataset for research purposes only
Not redistribute the dataset or any part of it
Cite the dataset in any publications resulting from its use
Adhere to the CC BY-NC 4.0 license terms

MIAM: A Multimodal Dataset for Enhancing Industrial Task Monitoring and Engagement Prediction

290 minutes of untrimmed industrial assembly and disassembly video across 22 sessions and 8 operators, captured with synchronized third-person RGB, top-view RGB and depth, and bimanual IMU streams, and annotated for action localization, object interaction, and real-time engagement state. The release links, per session: multi-view video clips, depth streams, left- and right-hand IMU traces, and frame-accurate annotations for actions, objects, and engagement. A 70/30 train/test split is provided for reproducible benchmarking.

Dataset viewer: the interactive viewer does not preview rows because the dataset is distributed as per-session .zip archives. The .zip files under train/ and test/ are files to browse and download. Each archive contains the full multimodal data for one session.

Targets and scope: industrial assembly and disassembly workflows performed by 8 operators across 22 sessions. Each session includes two camera views (third-person RGB, top-view RGB and depth) and bimanual IMU data, with annotations for action segments, object interactions, and engagement states. 70% of sessions form the training split and 30% the test split.

Repository layout

README.md                          this card
train/                             training split (70%)
  session_XX.zip                   per-session archive (video, depth, IMU, annotations)
test/                              test split (30%)
  session_XX.zip                   per-session archive (same structure as train)
docs/                              DATA_NOTES.md, ANNOTATION_GUIDE.md, scripts/
LICENSE.md, CITATION.cff           license and citation metadata

Each session_XX.zip extracts to:

session_XX/
  c1_clips/                        third-person RGB videos (MP4, H.264)
  c2_rgb_clips/                    top-view RGB videos
  c2_depth_clips/                  depth streams
  lh_imu_data.csv                  left-hand IMU (timestamped)
  rh_imu_data.csv                  right-hand IMU (timestamped)
  annotations.csv                  action + engagement labels

Files are distributed as one .zip per session to preserve directory structure and allow resumable downloads. Extract with unzip session_XX.zip.

Getting started

import pandas as pd
from pathlib import Path

session = Path("./extracted/session_01")

annotations = pd.read_csv(session / "annotations.csv")
lh_imu = pd.read_csv(session / "lh_imu_data.csv")
rh_imu = pd.read_csv(session / "rh_imu_data.csv")

print(f"Action segments: {len(annotations)}")
print(f"LH IMU samples: {len(lh_imu)}")
print(f"RH IMU samples: {len(rh_imu)}")

# annotations and docs only (small, fast)
hf download ArvindSihag/MIAM --repo-type dataset --include "docs/*" "README.md"

# a single session's archive
hf download ArvindSihag/MIAM --repo-type dataset --include "train/session_01.zip" --local-dir ./miam_data

# the full dataset
hf download ArvindSihag/MIAM --repo-type dataset --local-dir ./miam_data

Data description

Domain: industrial manufacturing and human-robot collaboration. Tasks: assembly and disassembly workflows. Annotation tool: VGG Video Annotator (VIA). Synchronization: Camera-1 hardware timestamp. Labels: actions, objects, engagement states. Video format: MP4 (H.264). IMU format: CSV (timestamped samples). Split: 70% train, 30% test.

Engagement prediction framework

MIAM supports multi-level behavioral analysis:

Action level: task recognition and temporal localization (RGB + depth)
Object level: interaction detection and tracking (RGB + IMU)
Engagement level: real-time engagement state prediction (all modalities)

License and usage policy

Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0).

Allowed: academic research, non-commercial projects, derivative works (non-commercial), publication with attribution.

Not allowed: commercial use, redistribution, re-identification of participants, removing license notices.

Usage rules:

Access requires manual approval.
Research purposes only.
Redistribution is not permitted.
Citation is mandatory in any publication.

Citation

If you use MIAM in your research, please cite:

@inproceedings{arvind2025multimodal,
  title     = {A Multimodal Dataset for Enhancing Industrial Task
               Monitoring and Engagement Prediction},
  author    = {Arvind and Mehta, Naval Kishore and Kumar, Himanshu
               and Saurav, Sumeet and Singh, Sanjay},
  booktitle = {Proceedings of the 20th ACM/IEEE International
               Conference on Human-Robot Interaction (HRI)},
  year      = {2025}
}

Related resources

Paper (arXiv): https://arxiv.org/abs/2501.05936 Code: https://github.com/arvindsihag Dataset: https://huggingface.co/datasets/ArvindSihag/MIAM

Acknowledgements

Developed at CSIR — Central Electronics Engineering Research Institute (CEERI), Pilani, and the Academy of Scientific and Innovative Research (AcSIR), India.

Contact

Arvind Sihag — arvind.ceeri24a@acsir.res.in

For queries, collaborations, or access issues, please reach out via email or open a discussion on this repository.

Downloads last month: 26

Total file size:

745 GB

Paper for ArvindSihag/MIAM

A Multimodal Dataset for Enhancing Industrial Task Monitoring and Engagement Prediction

Paper • 2501.05936 • Published Jan 10, 2025