Generalizing Visual Geometry Priors to Sparse Gaussian Occupancy Prediction

GPOcc leverages generalizable visual geometry priors (e.g., VGGT) and represents volumetric evidence as sparse 3D Gaussians for efficient monocular 3D occupancy prediction, with a training-free incremental fusion strategy for streaming inputs.

---

🔥 News

• Accepted to CVPR2026, code will be released before the conference.

---

📌 Citation

If you find this work useful, please consider citing:

@misc{zhou2026generalizingvisualgeometrypriors,
      title={Generalizing Visual Geometry Priors to Sparse Gaussian Occupancy Prediction}, 
      author={Changqing Zhou and Yueru Luo and Changhao Chen},
      year={2026},
      eprint={2602.21552},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2602.21552}, 
}