MMHAR-28: Human Action Recognition Across RGB, Depth, Thermal, and Event Modalities

Dataset Summary

MMHAR-28 is a multimodal human action recognition dataset designed for action classification across four sensing modalities:

RGB
Depth
Thermal
Event

The MMHAR-28 dataset contains 28 human action classes collected in two sessions. Session 1 focuses on single-person sports and exercise actions, while Session 2 focuses on two-person interaction activities.

Dataset Structure

The dataset is organized into predefined splits:

train
val
test

Samples are stored by modality. Typical modality folders include:

rgb_images
depth_images
thermal
event-streams

Data Instances

Each instance corresponds to one action sample and one label from the 28 action classes.

Example annotation format:

path/to/sample,label

Example paths:

data/train/session_1/sub_18/d_rgb/28/rgb_images,13
data/train/session_1/sub_7/d_rgb/26/depth_images,12
data/train/session_1/sub_33/thermal/9_1_0,8
data/train/session_1/sub_55/event-streams/15,7

Data Splits

The dataset provides predefined splits for:

training
validation
testing

Citation

If you use the MMHAR-28 dataset in your research, please cite our paper:

@ARTICLE{11447325,
  author={Rakhimzhanova, Tomiris and Kuzdeuov, Askat and Muratov, Artur and Varol, Huseyin Atakan},
  journal={IEEE Transactions on Biometrics, Behavior, and Identity Science}, 
  title={MMHAR-28: Human Action Recognition Across RGB, Thermal, Depth, and Event Modalities}, 
  year={2026},
  volume={},
  number={},
  pages={1-1},
  keywords={Videos;Cameras;Event detection;Thermal sensors;Sensors;Web sites;Video on demand;Three-dimensional displays;Software;Lighting;Human action recognition (HAR);multimodal learning;RGB;depth;thermal;event-based camera;multimodal dataset;video classification;deep learning},
  doi={10.1109/TBIOM.2026.3675639}}