Global SfM-style reconstruction

Glob3R

Global Structure-from-Motion with 3D foundation models, turning feed-forward geometric predictions into optimizable multi-view constraints.

Paper

Code

Demo videos

Qualitative results and process visualizations.

Large-scene reconstruction

Global alignment and dense geometry over a wide scene.

Compact-scene reconstruction

Stable camera motion and dense structure in a smaller capture.

Optimization process

Visualizing how initialization becomes a refined global reconstruction.

Novel-view synthesis

Refined camera poses improve downstream rendering quality.

Pipeline

Dense tracks, global association, and SfM refinement.

Given an ordered sequence or retrieval-based pseudo-sequence, Glob3R predicts local geometric priors and dense warps inside overlapping windows. Reliable tracks are merged into a global association graph before rotation averaging, translation averaging, bundle adjustment, and dense reconstruction.

Glob3R pipeline showing Pi3X geometric prediction, keyframe selection, dense matching, track generation, motion averaging, bundle adjustment, and dense reconstruction. — Pipeline rendered from the project figure. Local predictions are converted into sparse multi-view tracks, then optimized globally.

Citation

BibTeX

@article{glob3r2026,
  title={Glob3R: Global Structure-from-Motion with 3D Foundation Models},
  author={Deng, Junyuan and Li, Heng and Qiu, Kejie and Qiu, Lingteng and Peng, Rui and Shen, Weichao and Yuan, Weihao and Zhu, Siyu and Dong, Zilong and Tan, Ping},
  journal={arXiv preprint arXiv:2607.09225},
  year={2026}
}