Humans as Checkerboards: Calibrating Camera Motion Scale for World-Coordinate Human Mesh Recovery
(ICCV 2025)

Fengyuan Yang, Kerui Gu, Ha Linh Nguyen, Tze Ho Elden Tse,Angela Yao

[Paper], [Supp], [arXiv], [Project page]

1. Requirements

• Python 3.9
• PyTorch 1.11.0

2. Preparation

HAC calibrates monocular SLAM scale by leveraging human priors from HMR.
It is model-agnostic to both HMR and SLAM, compatibility with various HMR and SLAM approaches as long as you can export the following inputs:

2.1 Required Inputs

(A) HMR predictions

Provide a dict (or serialized file) like:

{
    "pred_rotmat": ...,  # Predicted SMPL rotation matrix 
    "pred_shape":  ...,  # Predicted SMPL body shape parameters
    "pred_trans":  ...   # Predicted camera translation
}

(B) SLAM camera estimates and 3D point cloud

Provide a dict (or serialized file) like:

{
    "traj":   ...,  # camera poses represented as [tx, ty, tz, qx, qy, qz, qw]
    "tstamp": ...,  # timestamps
    "disps":  ...   # disparity/depth proxy
}

together with SLAM‘s output 3D point cloud:

  points.ply

2.2 Reference Configuration

For SOTA comparison/evaluation, we used TRAM's setup:

• HMR: VIMO
• SLAM: Masked DroidSLAM

You can export the above inputs and points.ply from the source code and pass them directly to HAC.

3. Usage

• Test on EMDB 2

  python eval_emdb2.py >logs/emdb2.logs 2>&1 

  • Corresponding output logs can found at logs/emdb2.logs

• Test on EgoBody

  python eval_egobody.py >logs/egobody.logs 2>&1 

  • Corresponding output logs can found at logs/egobody.logs

Citation

If you find our paper or codes useful, please consider citing our paper:

@InProceedings{Yang_2025_ICCV,
    author    = {Yang, Fengyuan and Gu, Kerui and Nguyen, Ha Linh and Tse, Tze Ho Elden and Yao, Angela},
    title     = {Humans as Checkerboards: Calibrating Camera Motion Scale for World-Coordinate Human Mesh Recovery},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2025},
    pages     = {6069-6079}
}

Acknowledgments

Our codes are based on WHAM, TRAM, and SLAHMR and we really appreciate it.