Skip to content
/ PIP Public

A real-time system that captures physically correct human motion, joint torques, and ground reaction forces with only 6 inertial measurement units

License

Notifications You must be signed in to change notification settings

Xinyu-Yi/PIP

Repository files navigation

PIP

Code for our CVPR 2022 paper "Physical Inertial Poser (PIP): Physics-aware Real-time Human Motion Tracking from Sparse Inertial Sensors". This repository contains the system implementation and evaluation. See Project Page.

1

Usage

Install dependencies

We use python 3.7.6. You should install the newest pytorch chumpy vctoolkit open3d pybullet qpsolvers cvxopt.

You also need to compile and install rbdl with python bindings. Also install the urdf reader addon. This library is easy to compile on Linux. For Windows, you need to rewrite some codes and the CMakeLists. We have only tested our system on Windows.

If the newest vctoolkit reports errors, please use vctoolkit==0.1.5.39.

Installing pytorch with CUDA is recommended but not mandatory. During evaluation, the motion prediction can run at ~120fps on CPU, but computing the errors may be very slow without CUDA.

If you have configured TransPose, just use its environment and install the missing packages including the rbdl.

Prepare SMPL body model

  1. Download SMPL model from here. You should click SMPL for Python and download the version 1.0.0 for Python 2.7 (10 shape PCs). Then unzip it.
  2. In config.py, set paths.smpl_file to the model path.

If you have configured TransPose, just copy its settings here.

Prepare physics body model

  1. Download the physics body model from here and unzip it.
  2. In config.py, set paths.physics_model_file to the body model path.
  3. In config.py, set paths.plane_file to plane.urdf. Please put plane.obj next to it.

The physics model and the ground plane are modified from physcap.

Prepare pre-trained network weights

  1. Download weights from here.
  2. In config.py, set paths.weights_file to the weights path.

Prepare test datasets

  1. Download DIP-IMU dataset from here. We use the raw (unnormalized) data.
  2. Download TotalCapture dataset from here. You need to download the real world position and orientation under Vicon Groundtruth in the website and unzip them. The ground-truth SMPL poses used in our evaluation are provided by the DIP authors. So you may also need to contact the DIP authors for them.
  3. In config.py, set paths.raw_dipimu_dir to the DIP-IMU dataset path; set paths.raw_totalcapture_dip_dir to the TotalCapture SMPL poses (from DIP authors) path; and set paths.raw_totalcapture_official_dir to the TotalCapture official gt path. Please refer to the comments in the codes for more details.

If you have configured TransPose, just copy its settings here. Remember: you need to rerun the preprocess.py as the preprocessing of TotalCapture dataset has been changed to remove the acceleration bias.

Run the evaluation

You should preprocess the datasets before evaluation:

python preprocess.py
python evaluate.py

The pose/translation evaluation results for DIP-IMU and TotalCapture test datasets will be printed/drawn.

Live Demo

The live demo codes are on the livedemo branch. Please checkout this branch.

About the codes

The authors are too busy to clean up/rewrite the codes. Here are some useful tips:

  • In dynamics.py, there are many disabled options for the physics optimization. You can try different combinations of the energy terms by enabling the corresponding terms.

  • In Line ~44 in net.py:

    self.dynamics_optimizer = PhysicsOptimizer(debug=False)

    set debug=True to visualize the estimated motions using pybullet. You may need to clean the cached results and rerun the evaluate.py. (e.g., set flush_cache=True in evaluate() and rerun.)

  • In Line ~244 in dynamics.py:

    if False:   # visualize GRF (no smoothing)
        p.removeAllUserDebugItems()
        for point, force in zip(collision_points, GRF.reshape(-1, 3)):
            p.addUserDebugLine(point, point + force * 1e-2, [1, 0, 0])

    Enabling this to visualize the ground reaction force. (You also need to set debug=True as stated above.) Note that rendering the force lines can be very slow in pybullet.

  • The hyperparameters for the physics optimization are all in physics_parameters.json. If you set debug=True, you can adjust these parameters interactively in the pybullet window.

Citation

If you find the project helpful, please consider citing us:

@InProceedings{PIPCVPR2022,
  author = {Yi, Xinyu and Zhou, Yuxiao and Habermann, Marc and Shimada, Soshi and Golyanik, Vladislav and Theobalt, Christian and Xu, Feng},
  title = {Physical Inertial Poser (PIP): Physics-aware Real-time Human Motion Tracking from Sparse Inertial Sensors},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  month = {June},
  year = {2022}
}

About

A real-time system that captures physically correct human motion, joint torques, and ground reaction forces with only 6 inertial measurement units

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages