Jiayi Liu¹  Jiaming Zhou¹  Ke Ye¹  Kun-Yu Lin²  Allan Wang³  Junwei Liang^1,4 
1HKUST(GZ) ²HKU ³Miraikan ⁴HKUST

* Cyan highlights occlusion-induced gaps; red indicates ID switches; green shows ego-centric perspective distortions.

* Dashed: First person view derived; Solid: bird's eye view derived.

EgoTraj-Bench is a real-world benchmark for robust trajectory prediction from ego-centric noisy observations. It grounds noisy first-person visual histories in clean bird’s-eye-view future trajectories, explicitly modeling real-world perceptual artifacts such as occlusions, ID switches, and tracking drift.

BiFlow, our dual-stream flow matching model, jointly denoises noisy ego-centric histories and predicts future trajectories via a shared latent representation, enhanced by an EgoAnchor mechanism for robust intent modeling.

• Release benchmark code and repository structure. (by Mar. 2026)
• Release benchmark dataset and download instructions. (by Mar. 2026)
• Release baseline models and evaluation scripts. (by Mar. 2026)
• Add detailed documentation for data format, metrics, and leaderboard. (by Apr. 2026)
• Add examples and tutorials for using EgoTraj-Bench. (by Apr. 2026)

If you encounter any problems or have questions about EgoTraj-Bench, please feel free to open an issue on the GitHub repo.

If you find our work helpful, please consider starring this repo 🌟 and cite:

@article{liu2025egotraj,
    title   =   {EgoTraj-Bench: Towards Robust Trajectory Prediction Under Ego-view Noisy Observations},
    author  =   {Liu, Jiayi and Zhou, Jiaming and Ye, Ke and Lin, Kun-Yu and Wang, Allan and Liang, Junwei},
    journal =   {arXiv preprint arXiv:2510.00405},
    year    =   {2025}
}

License will be updated.

Project page template is based on Nerfies.