WiFlow: Estimating Optical Flow using WiFi Channel State Information

WiFlow: Estimating Optical Flow using WiFi Channel State Information
Thomas Weigel, Simon Kiefhaber, Fabian Portner, Matthias Hollick, Simone Schaub-Meyer
TU Darmstadt · TU Delft · hessian.AI

[Paper] [Dataset] [Project Page]

Requirements

The code has been tested with Python 3.10 and PyTorch 2.0.

git clone https://github.com/visinf/wiflow.git
cd wiflow
pip install -e .

Or install directly from the Git URL:

pip install "wiflow @ git+https://github.com/visinf/wiflow.git"

Required Data

Note: The WiFlow dataset will be released publicly soon. Please check the project page for updates.

Once available, download the dataset and place it under dataset/. The dataset provides three aligned variants — sideview, birdview, and birdviewplus — with synchronized CSI and pseudo ground truth optical flow. Dataset configs are provided in dataset_configs/; point the config's dataset_root to your download location.

Pseudo Ground Truth Generation

If you want to generate pseudo-GT flow from your own camera frames, use the notebook tools/generate_psudo_gt.ipynb. It runs an ensemble of five optical flow models (rpknet, ms_raft_p, sea_raft_m, memflow, dpflow) via PTLFlow.

Training

WiFlowSimple:

python train_recurrent.py

WiFlowRoI (two-step — pretrain mask, then full model):

python train_mask.py
python train_roi.py

WiFlowCombo:

python train_recurrent.py   # train flow branch
python train_mask.py        # train mask branch
# then combine weights via WIFlow/WIFlow_combined.py

Training logs and checkpoints are written to runs/.

Evaluation

python evaluation.py

We report EPE (all pixels), EPE_M (moving pixels), EPE_S (static pixels), and EPE_A (amplified, power 4). See the paper for full results.

Benchmarking

python benchmarking/main.py

Citation

@inproceedings{weigel2025wiflow,
  title     = {WiFlow: Estimating Optical Flow using WiFi Channel State Information},
  author    = {Weigel, Thomas and Kiefhaber, Simon and Portner, Fabian and Hollick, Matthias and Schaub-Meyer, Simone},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026},
}