Code for our SIGGRAPH 2023 paper "EgoLocate: Real-time Motion Capture, Localization, and Mapping with Sparse Body-mounted Sensors". See Project Page.

Authors: Xinyu Yi, Yuxiao Zhou, Marc Habermann, Vladislav Golyanik, Shaohua Pan, Christian Theobalt, Feng Xu

EgoLocate estimates accurate human pose, localization, and reconstructs the scene in sparse 3D points from 6 IMUs and a head-mounted camera. Importantly, EgoLocate does not rely on pre-scanning the scene, and runs in real time on CPU.

The SLAM library is implemented in C++ based on ORB-SLAM3. Please follow the instructions in the slam/ folder to build the library.

The mocap module is implemented in Python based on our previous works PIP and TransPose. Please follow the instructions in the mocap/ folder to prepare the dependencies.

We provide a minimal 30-second live example for the users to examine whether the system is configured correctly. We suggest compiling the slam library with use_viewer=true. Then, run:

python test_live_recording.py

The system will first perform the curve initialization. When the initialization is finished, calibration matrices SMC0 and RHC will be printed (need VerboseLevel=4 in 3.yaml, the default setting). After a period of walking, when the subject go back to the starting position, the system will perform a loop closure optimization.

• TotalCapture dataset

  To evaluate our method on TotalCapture dataset, we synthesized monocular videos using a virtual first-person camera in 3 virtual scenes of different sizes. Please first contact the TotalCapture authors to get the permission to the dataset. Then, download the synthesized videos, data, and further instructions here. (coming soon...)

• HPS dataset

  We are still collecting scripts used for HPS dataset...

• To run the method for one sequence in the TotalCapture dataset, execute:

  python test_totalcapture.py --run <seq_idx> <n>

  This will test the <seq_idx>th sequence (0~44) and save the result as seq_<seq_idx>_<n>.pt. It includes the estimated pose and translation. If you want to visually compare the estimated motion with TransPose/PIP in real time during running, add --visualize (black=ground truth; green=ours; white=PIP).

  Note: you may also set use_viewer=true when configuring the SLAM library to have a full visualization. For more debug information, modify VerboseLevel in the yaml configuration file. TotalCapture use 0.yaml.

• To run the method on the whole TotalCapture dataset for a full test, run:

  ./test.sh

  This will run all 45 sequences 9 times and will take about 9 hours.

• After running the method on the whole dataset, get the translation error by:

  python test_totalcapture.py --evaluate <motion_type>

  This will read the saved results and compute the absolute translation errors and standard deviations. You can use any one of acting, freestyle, rom, walking, all for the <motion_type>. This will evaluate the specific type of motions for TotalCapture dataset.

Testing on a laptop with Intel(R) Core(TM) i7-12700H CPU, our method runs at ~70fps. The method runs purely on CPU in real time.

Note that if you have a high-performance CPU and the method runs much faster than 60fps, you may need to set use_clock = True in Line 11 in egolocate.py. This will force the mocap to run at 60fps, giving enough time for the back-end mapping.

Authors: Due to the size of this project, the instructions provided may not be comprehensive. Should you encounter any difficulties or notice any important bugs, please leave an issue and the author will promptly update the code.

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

@article{EgoLocate2023,
    author = {Yi, Xinyu and Zhou, Yuxiao and Habermann, Marc and Golyanik, Vladislav and Pan, Shaohua and Theobalt, Christian and Xu, Feng},
    title = {EgoLocate: Real-time Motion Capture, Localization, and Mapping with Sparse Body-mounted Sensors},
    journal={ACM Transactions on Graphics (TOG)},
    year = {2023},
    volume = {42},
    number = {4},
    numpages = {17},
    articleno = {76},
    publisher = {ACM}
}