Official code for our ICCV25 paper “MagShield: Towards Better Robustness in Sparse Inertial Motion Capture Under Magnetic Disturbances”. See project page.

Our algorithm requires minimal dependencies: Python 3.8 (strictly required, as ESKF relies on pybind11 compilation) and PyTorch 1.13.1 (our tested version).

For full integration with existing inertial posers (PNP and DynaIP), you'll need to install additional dependencies. We recommend you configure the environment dependencies in the following order:

1. Create a Python 3.8 environment
2. Configure PNP (most complex setup)
3. Configure DynaIP

Download SMPL model from here. You should download the version 1.0.0 for Python 2.7 (10 shape PCs). Rename and put the male model file into models/SMPL_male.pkl.

1. Download PNP weights from here, put it into inertial_poser/PNP/weights/weights.pt
2. Download DynaIP weights from here, put it into inertial_poser/DynaIP/weights/weights.pt

1. Download MagIMU dataset from here
2. Put the unziped data into datasets/MagIMU. The structure should be like:

datasets
└── MagIMU
    ├── gt
    	├── xxx.pt
    ├── raw_measurements
    	├── xxx.pt
    ├── readings_noitom
    	├── xxx.pt
    └── align.py

The evaluation configurations are saved in exp. For each inertial poser, we provide three settings: ESKF9, ESKF9+detector, ESKF9+detector+corrector.

To get the mocap result, run:

python predict.py -c "exp/{INERTIAL_POSER_NAME}/{CONFIG}"

To get the evaluate results, run:

python eval.py

With a successful run, the quantitative results should be as follows:

for PNP:

for DynaIP:

These results are not entirely consistent with those reported in the original paper. This is due to two reasons: 1) We fixed a minor bug in the timestamp down-sampling process; 2) We added an 2d embedding to the output layer of the corrector.