Coherent4D

From Where to How: Continuous 4D Interaction Forecasting from Egocentric Video

Qiaohui Chu^1,2, Haoyu Zhang^1,2, Meng Liu^3,, Haoxiang Shi^1,2, Dongmei Jiang², Liqiang Nie^1,

¹ Harbin Institute of Technology (Shenzhen) · ² Pengcheng Laboratory · ³ Shandong University
^* Corresponding authors: Meng Liu and Liqiang Nie.

Paper · Project page

Coherent4D pairs future 3D interaction locations with temporally aligned full-body poses in a shared coordinate system. The interaction and WHAM intermediates were generated by the authors.

Domain	Future events	Train	Validation	Test
Cooking	10	136,079	15,435	14,527
Health	5	21,532	1,899	4,167
Bike Repair	4	35,987	3,150	1,052
Total	—	193,598	20,484	19,746

Each domain contains train.pkl.gz, val.pkl.gz, test.pkl.gz, and corresponding take lists under splits/. These archives contain annotations, not source videos, model weights, SMPL templates, or joint regressors.

Load annotations

Download the annotations with the Hugging Face CLI:

hf download Qiu0710/Coherent4D --repo-type dataset --local-dir Coherent4D
cd Coherent4D

import gzip
import pickle
import numpy as np

with gzip.open("bike_repair/test.pkl.gz", "rb") as f:
    data = pickle.load(f)

sample = data["dataset"][0]
paired = np.asarray(sample["target_mask"], dtype=bool) & (
    sample["target_pose_mask"].astype(bool)
)
locations = np.asarray([
    event["location_norm"]
    for event, valid in zip(sample["target_interactions"], paired)
    if valid
], dtype=np.float32)
joints_m = sample["target_joints_f0_m"][paired]

Only load trusted pickle files. Verify downloads with sha256sum -c SHA256SUMS.

• Locations use a fixed right-hand SMPL mesh landmark (vertex 5777), divided by 5 m and clipped to [-1, 1].
• Root translations and 19-joint positions remain in meters in the same local frame. Body rotations contain 23 parent-relative 6D rotations.
• time_s is measured within a take, not calendar time. time_rel_s starts at the first future event.
• Use validity masks to exclude padding. Some samples have empty observation histories.
• The publication copy removes machine-specific configuration paths. Take identifiers, shape parameters, and motion labels are retained for reproducibility; this is not a fully anonymized release.

Source and acknowledgments

These annotations were constructed from Ego-Exo4D, using FIction, WHAM, and SMPL. Source videos and body-model assets are obtained separately under their respective terms. We thank the Ego-Exo4D team and participants.

@inproceedings{grauman2024ego,
  title={Ego-exo4d: Understanding skilled human activity from first- and third-person perspectives},
  author={Grauman, Kristen and Westbury, Andrew and Torresani, Lorenzo and Kitani, Kris and Malik, Jitendra and Afouras, Triantafyllos and Ashutosh, Kumar and Baiyya, Vijay and Bansal, Siddhant and Boote, Bikram and others},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={19383--19400},
  year={2024}
}