Ego-Exo4D-HM/npz-datasets

Ego-Exo4D Human Meshes (Ego-Exo4D-HM)

4D human motion reconstructions for Ego-Exo4D captures.

Paper · Project page · Code

Ego-Exo4D provides synchronized egocentric and multi-view exocentric video, but ships with only sparse 3D human pose annotations. Ego-Exo4D-HM fills that gap with dense, per-frame SMPL-H body and hand motion for the camera wearer in each take, reconstructed in Ego-Exo4D's metric world frame.

The motion is recovered with a multi-view pipeline built on SLAHMR. Person detection is guided by the Aria head trajectory. Body keypoints come from ViTPose and hand keypoints from HaMeR. The keypoints are triangulated across the calibrated exocentric cameras, and SMPL-H is fit to the result.

Dataset at a glance

Takes	2,649
Frames	11.3M (30 fps)
Reconstructed motion	104.6 hours
Video covered	523 hours (4 exocentric + 1 egocentric view per take)
Size	~48.5 GB (one .npz per take)

Activities: basketball (817), dance (587), cooking (360), climbing (348), bike repair (254), health (123), soccer (92), music (68)

Institutions: uniandes (852), sfu (387), unc (371), upenn (195), iiith (194), minnesota (161), georgiatech (143), indiana (120), cmu (81), utokyo (74), fair (47), nus (24)

Takes were filtered for reconstruction quality before release. See Section 4.1 of the paper for the criteria.

Structure

One folder per Ego-Exo4D take, named with the take's Ego-Exo4D take_name:

<take_name>/points_triangulated_world_results_merged.npz

Each file merges the optimizer's per-chunk results over the full take. In the table below, T is the number of frames.

Key	Shape	Description
trans	(1, T, 3)	Global root translation, in Ego-Exo4D world coordinates (meters)
root_orient	(1, T, 3)	Global root orientation (axis-angle)
pose_body	(1, T, 63)	SMPL-H body pose (21 joints × 3, axis-angle)
hand_pose	(1, T, 90)	Hand pose, both hands (2 × 15 joints × 3)
latent_pose	(1, T, 32)	VPoser latent body pose
betas_per_frame	(1, T, 16)	Body shape. Constant within each optimization chunk and repeated for every frame
joints3d	(1, T, 67, 3)	3D keypoints (world frame)
joints2d	(4, T, 67, 2)	3D keypoints reprojected into the 4 exocentric views
cam_R, cam_t	(5, T, 3, 3) / (5, T, 3)	Camera rotation and translation (Ego-Exo4D calibration, not optimized)
intrins, cam_dist	(5, T, 3, 3) / (5, T, 5)	Camera intrinsics and distortion coefficients
valid	(1, T) int8	0 for frames whose optimization chunk hit a NaN/Inf loss term, 1 otherwise
chunk_ranges	(n_chunks, 2)	(start, end) frame indices of each optimization chunk

67-keypoint layout (joints3d, joints2d): 0–24 are OpenPose body keypoints, 25–45 are the left hand (21 keypoints), and 46–66 are the right hand (21 keypoints).

Usage

Download one take:

pip install -U "huggingface_hub[cli]"
hf download Ego-Exo4D-HM/npz-datasets --repo-type dataset \
    --include "cmu_bike02_4/*" --local-dir data_filtered

Download the whole dataset. It is safe to re-run, and files that are already downloaded are skipped. If the download stalls, retry with --max-workers 4:

hf download Ego-Exo4D-HM/npz-datasets --repo-type dataset --local-dir data_filtered

Load a take:

import numpy as np

d = np.load("data_filtered/cmu_bike02_4/points_triangulated_world_results_merged.npz")
print(d["trans"].shape)    # (1, T, 3)
print(d["joints3d"].shape) # (1, T, 67, 3)
good = d["valid"][0].astype(bool)  # mask out frames from unstable chunks

To get meshes, pass the parameters through an SMPL-H body model. To render them over the original video, use scripts/run_mesh_vis_hands_egoexo.py from the code repository. Rendering needs the raw Ego-Exo4D videos.

Evaluation

Measured against Ego-Exo4D's ground-truth 3D keypoint annotations, with no alignment (global MPJPE), the reconstructions reach 56.21 mm on body keypoints (845 annotated takes) and 51.59 mm on hand keypoints (190 annotated takes).

License

Released under CC BY 4.0.

Citation

@misc{maddukuri2026egoexo4dhumanmeshesdataset,
      title={Ego-Exo4D Human Meshes Dataset: 4D Human Motion Reconstruction for Ego-Exo Captures},
      author={Abhiram Maddukuri and Georgios Pavlakos},
      year={2026},
      eprint={2609.30187},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2609.30187},
}