R³: 3D Reconstruction via Relative Regression
Model Description
R³ is a feed-forward model for long-video 3D reconstruction via confidence-weighted relative pose regression.
Built on top of the Depth Anything 3 (DA3) backbone, R³ predicts pairwise relative camera poses and dense geometry from arbitrarily long video streams. Instead of directly regressing all cameras in a global coordinate frame, R³ estimates confidence-weighted relative transformations between image pairs and assembles a globally consistent trajectory downstream.
With approximately 372M parameters, R³ achieves competitive or state-of-the-art performance on camera pose estimation and dense reconstruction benchmarks while remaining significantly smaller than recent billion-parameter alternatives.
License
This model is a derivative work based on the Depth Anything 3 (DA3) model.
The original DA3 model is released under the CC BY-NC 4.0 license. Accordingly, this model is also distributed under the CC BY-NC 4.0 license and may only be used in accordance with the terms of that license.
Commercial use is not permitted without obtaining appropriate permissions from the original rights holders.
Model Details
- Model Type: 3D Reconstruction / Visual Localization
- Backbone: Depth Anything 3 (DA3)
- Parameters: ~372M
- Task: Relative Pose Regression and Dense 3D Reconstruction
Citation
If you use this model in your research, please cite the R³ paper.
@misc{xu2026r3,
title={$R^3$: 3D Reconstruction via Relative Regression},
author={Congrong Xu and Huachen Gao and Xingyu Chen and Yuliang Xiu and Jun Gao and Anpei Chen},
year={2026},
}