ON3R: Sparse–View Localization via Online Neural 3D Regression

Ludvig Dillén, Magnus Oskarsson, Viktor Larsson

Lund University
CVPR 2026

The official GitHub code repository for ON3R by Ludvig Dillén is available at github.com/LudvigDillen/ON3R .

The absolute pose estimation process for a new query (bottom left). We supervise it using reprojection errors and monodepth in the reference views (top). After reconstructing an initial point cloud, we estimate the absolute query pose with P3P-RANSAC and then refine it with a lightweight bundle adjustment (this is where things often “snap into place”). The reconstructed point cloud and the ground truth poses are shown in the bottom right. Note, the blue lines represent the reprojection residuals for each view. All videos run in real time.

Abstract

We present ON3R, an online-trained neural regressor addressing sparse-view structureless localization, where database images have limited visual overlap and no prebuilt 3D map. Given any sparse matches between a query and a $K$-tuple of posed database views, ON3R predicts 3D coordinates for matched query keypoints, supervised by database reprojection residuals and a monocular depth prior. Afterwards, the absolute pose of the query is estimated via P3P-RANSAC and refined with lightweight bundle adjustment. Across MegaDepth, Cambridge Landmarks, and a sparsified version of Aachen Day-Night, ON3R outperforms existing methods. ON3R is particularly effective when the data is extremely sparse -- we focus on $K\!\leq\!10$ database images.

Example 1

All-in-one video (top: reference views, bottom-left: query view with, bottom-right: reconstructed point cloud and ground truth poses).

Query image

Reference image 0

Reference image 1

Visualized reprojection errors (with ground truth pose) in query view.

Visualized reprojection errors in reference view 0.

Visualized reprojection errors in reference view 1.

Reconstruction during the learning/optimization process.

Example 2

All-in-one video (top: reference views, bottom-left: query view with, bottom-right: reconstructed point cloud and ground truth poses).

Query image

Reference image 0

Reference image 1

Visualized reprojection errors (with ground truth pose) in query view.

Visualized reprojection errors in reference view 0.

Visualized reprojection errors in reference view 1.

Reconstruction during the learning/optimization process.

Example 3

All-in-one video (top: reference views, bottom-left: query view with, bottom-right: reconstructed point cloud and ground truth poses).

Query image

Reference image 0

Reference image 1

Visualized reprojection errors (with ground truth pose) in query view.

Visualized reprojection errors in reference view 0.

Visualized reprojection errors in reference view 1.

Reconstruction during the learning/optimization process.

BibTeX

          @inproceedings{dillen2026on3r,
            title={{Sparse–View Localization via Online Neural 3D Regression}},
            author={Dillén, Ludvig and Oskarsson, Magnus and Larsson, Viktor},
            booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition},
            year={2026}
            }