CAM3R — camera-agnostic two-view 3D reconstruction

Weights for CAM3R, a two-view 3D reconstruction network that takes an image pair from arbitrary, possibly different camera models — pinhole, fisheye, or 360° panoramic — with no intrinsics given at test time.

• Paper: arXiv:2603.22631
• Code: github.com/nam1410/cam3r
• Project page: nam1410.github.io/cam3r

Usage

See github.com/nam1410/cam3r for installation, the inference scripts, and the expected input format.

from huggingface_hub import hf_hub_download
path = hf_hub_download("gnam8991/cam3r", "cam3r_release_weights.pth")

Citation

@inproceedings{guruprasad2026cam3r,
  title         = {CAM3R: Camera-Agnostic Model for 3D Reconstruction},
  author        = {Guruprasad, Namitha and Yadav, Abhay and Chellappa, Rama and Peng, Cheng},
  eprint        = {2603.22631},
  archivePrefix = {arXiv},
  booktitle     = {European Conference on Computer Vision (ECCV)},
  year          = {2026}
}

License

CC BY-NC-SA 4.0, inherited from DUSt3R / CroCo, on which this code is based. Non-commercial use only.