EgoTraj

Real-world egocentric human trajectories with head pose, eye gaze, video and scene annotations.

ECCV 2026 · Paper · Paper page · Code & loader · Preprint with appendix

At a glance

Sessions	Participants	Recording	RGB frames	Rate	Device
75	75	10.7 h	1.15M	30 Hz	Meta Quest Pro

Participants walked self-chosen routes through real urban streets, sidewalks and crosswalks.

What each session contains

Stream	Details
Head pose	6-DoF position and orientation, plus linear and angular velocity
Eye gaze	3D gaze origin and direction for every frame
Video	Egocentric RGB, with faces and licence plates blurred
Scene annotations	One per second: context, gaze target and inferred intent

Every stream is time-aligned at 30 Hz. Here is one session in the EgoViz viewer:

And two of the per-second scene annotations:

Files

File	Contents
egotraj_dataset.h5	Pose, gaze and video pointers; one group per session
egotraj_annotations.json	Scene annotations, keyed by session and pose_idx
egotraj_videos_part1.zip … part5.zip	Video; download all five parts

Quick start

pip install huggingface_hub git+https://github.com/yehiahmad/EgoTraj

from huggingface_hub import hf_hub_download
from egotraj import list_sessions, load_session, load_annotations

h5 = hf_hub_download("AhmadYehia/EgoTraj", "egotraj_dataset.h5", repo_type="dataset")
s = load_session(h5, list_sessions(h5)[0])

s.position          # (N, 3) metres, session-relative, Y is up
s.gaze_direction    # (N, 3) unit gaze vectors
s.timestamp         # (N,)   epoch seconds

Benchmark

The paper's task is egocentric trajectory prediction: observe 1.5 s, predict the next 3.5 s, scored by ADE, FDE and head-orientation error.

License and ethics

The data is released under CC BY-NC 4.0 for non-commercial use; the loader code on GitHub is MIT. Collection was approved by an Institutional Review Board, every participant gave written informed consent, and faces and licence plates are blurred. Do not attempt to identify anyone in the data.

Citation

@inproceedings{yehia2026egotraj,
  title     = {EgoTraj: Real-World Egocentric Human Trajectory Dataset for Multimodal Prediction},
  author    = {Yehia, Ahmad and Mohamed, Abduallah and Wang, Tianyi and Byeon, Jiseop and Qian, Kun and Jiao, Junfeng and Claudel, Christian},
  booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
  year      = {2026}
}