Reproduce “EgoTraj: Real-World Egocentric Human Trajectory Dataset for Multimodal Prediction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ahmad Yehia, Abduallah Mohamed, Tianyi Wang, Jiseop Byeon, Kun Qian, Junfeng Jiao, Christian Claudel
EgoTraj is presented as a large-scale, real-world egocentric dataset for human trajectory prediction, integrating synchronized 6DoF head pose, 3D gaze, RGB video, and scene annotations collected from 75 diverse participants in urban settings. Benchmarking on EgoTraj reveals that multimodal models, particularly those incorporating gaze and scene context, enhance prediction accuracy, with the EgoCast model achieving an Average Displacement Error of 0.16 meters and a Final Displacement Error of 0.28 meters.
Key insights
- Multimodal cues, especially gaze and scene context, provide significant anticipatory information for egocentric trajectory prediction, improving accuracy beyond purely kinematic observations.
- Fine-grained representations of social context, such as body poses of nearby people, are more informative than simpler bounding box or center point representations for modeling social interactions.
- Learned temporal models consistently outperform classical kinematic baselines for both position and head orientation prediction, with cross-stream attention mechanisms proving effective in egocentric contexts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.