M3-CFR

M3-CFR is a bi-static mmWave MIMO channel frequency response (CFR) dataset for fine-grained micro-gesture recognition under domain shifts. The dataset and paper target communication-native sensing in integrated sensing and communications (ISAC), using CFR captured from an 8 x 8 fully digital mmWave MIMO testbed rather than radar-specific sensing hardware.

This repository currently contains:

• a PyTorch dataset loader and windowing utility in m3cfr_datagen.py
• a VGG-16 baseline training script in train_vgg16_v2.py
• the accompanying paper describing the dataset and benchmark task

Dataset download: https://huggingface.co/datasets/foysalhaque/M3-CFR/tree/main

Citation

If you use M3-CFR, please cite the paper:

@article{haque2026m3cfr,
  title={M3-CFR: A Domain-Adaptive Bi-Static mmWave MIMO CFR Dataset for Micro-Gesture Recognition},
  author={Haque, Khandaker Foysal and Meneghello, Francesca and Restuccia, Francesco},
  journal={IEEE Sensors Letters},
  year={2026}
}

Overview

M3-CFR dataset contains:

• 10 micro-gesture classes
• 3 subjects
• 3 indoor environments
• 3,000 labeled gesture instances
• about 1.5 million CFR frames in total
• 8 x 8 MIMO CFR measurements at 58 GHz with 1 GHz bandwidth
• 1024 OFDM subchannels per CFR snapshot

Micro-gestures

The dataset covers the following gesture categories:

• thumbs up
• thumbs down
• victory
• okay
• pointing
• pinch
• tap fingers
• fingers crossed
• fist
• half open palm

Expected Dataset Layout

The dataset loader expects a directory structure of the form:

<root_dir>/
|-- <env>/
|   `-- <subject>/
|       |-- Finger-Crossed/
|       |-- Fist/
|       |-- half-palm/
|       |-- okay/
|       |-- pinch/
|       |-- pointing/
|       |-- tap-fingers/
|       |-- thumbs-down/
|       |-- thumbs-up/
|       `-- victory/

Each gesture folder contains .mat files. The loader assumes each file stores one CFR tensor with shape:

1024 x 8 x 8 x frames

The code automatically picks the first non-metadata variable from each .mat file.

Data Loader

m3cfr_datagen.py provides:

• M3CFRDataset, a PyTorch Dataset for loading gesture instances
• temporal window extraction with configurable window_size and stride
• conversion from complex CFR to concatenated real/imaginary features
• per-sample normalization
• create_train_test_split() for a simple random train/test split

For each sample, the loader:

1. loads a CFR tensor from a .mat file
2. extracts a temporal window
3. reshapes 8 x 8 MIMO links into 64 spatial streams
4. concatenates real and imaginary parts
5. returns a tensor of shape Tw x 1024 x 128

Training Baseline

train_vgg16_v2.py is an example training script built around torchvision.models.vgg16_bn.

It currently:

• trains on one selected environment and one selected subject at a time
• uses window_size=10 and stride=5
• performs a random 80/20 train/test split
• runs a one-batch overfit sanity check before full training
• saves the best checkpoint as best_vgg16_bn_<env>_<subject>.pth

Run Training

After downloading and extracting the dataset into the expected folder structure:

python train_vgg16_v2.py

For any question or query, please contact Foysal Haque (haque.k@northeastern.edu)