@misc{zhang2025emergentextremeviewgeometry3d,
      title={Emergent Extreme-View Geometry in 3D Foundation Models}, 
      author={Yiwen Zhang and Joseph Tung and Ruojin Cai and David Fouhey and Hadar Averbuch-Elor},
      year={2025},
      eprint={2511.22686},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2511.22686}, 
}

1. Clone the repository (with submodules):

   git clone --recurse-submodules https://github.com/theREALevan/extreme-view-3dfm.git
   cd extreme-view-3dfm

2. Install dependencies: Go into the corresponding model folder under models/ and follow that repo’s setup.

   • VGGT: models/vggt — see VGGT Quick Start.
   • π³ (Pi3): models/pi3 — see Pi3 Quick Start.
   • World-Mirror (WM): models/worldmirror — see Dependencies and Installation.

1. In scripts/eval.sh, set BASE_DIR to your MegaUnScene data root.

2. Set MODEL to vggt, pi3, or wm. Set NO_CKPT=1 to use the pre-trained model only; otherwise the fine-tuned checkpoint is used.

3. Run:

   ./scripts/eval.sh

For evaluations on monocular depth, multiview pose estimation, and dense reconstruction (including on UnSceneRecon), please refer to the generalization evaluation GitHub repo.

For MegaUnScene information and download instructions, please refer to the HuggingFace dataset page.