GEOMETRIC REFINEMENT FOR CAMERA POSES

PosePlaner: Denoising Any Feed-Forward Pose Predictor
from Pairwise Planar Geometry

Yihang Chen^1,2Bin Tan^2,‡Yujun Shen²Yinghao Xu^1,‡Nan Xue²

¹ HKUST, Hong Kong, China² Robbyant, Ant Group, China

‡ Corresponding authors

Refine an existing camera pose with matched planes in one flow-matching step.
Train on image pairs; transfer the same refiner to long sequences.

Turn geometric evidence into a correction for learned camera poses.

Learning pose refinement from matched planes

PosePlaner conditions a diffusion model on matched planes and an initial camera pose to learn pose corrections. A correct relative pose should align corresponding planes across views. Trained with flow matching, the refiner learns a pose correction in a single denoising step.

Illustrative GT, not a model input. One of multiple matched plane pairs shown.

Refinement reduces plane disagreement

Before refinement
After refinement

Normal disagreement

Median normal residual ‖Δn‖₂
Dataset	Before refinement	After refinement
ScanNet1500	0.200	0.084
7Scenes	0.156	0.076

Position disagreement

Median scene-normalized offset residual |Δd|
Dataset	Before refinement	After refinement
ScanNet1500	0.072	0.064
7Scenes	0.057	0.053

Median residuals from the paper’s LoFTR experiments; lower is better. Left: normal difference. Right: scene-normalized plane offset. These are measured results, separate from the schematic above.

METHOD

One refiner, multiple geometric hypotheses

A flow-matching Transformer predicts candidate pose corrections in one Euler step. A learned score head fuses the candidates using planar consistency.

PosePlaner architecture: plane prediction and correspondence matching produce geometry tokens that condition a multi-query diffusion refiner and score-based hypothesis fusion. — One refiner, conditioned on the input pose and matched plane pairs.

Match local planes

Predict planar primitives in each camera’s local frame. Convert pixel matches into patch-level correspondences and compute plane-warping residuals.

Denoise and fuse

A flow-matching Transformer produces multiple pose hypotheses in one Euler step. A learned score head combines them using geometric consistency.

Recover two-view translation

With the refined rotation fixed, recover translation using a geometric optimization initialized from the upstream estimate.

Correct long trajectories

Refine pairwise rotations in the trajectory tail. Rotation averaging and global positioning use a shared reconstruction backend.

Results

A geometric correction for two-view poses and streaming trajectories.

Seven upstream models on ScanNet1500 and zero-shot 7Scenes. Rotation AUC@10° (%).

Dataset

ScanNet1500 · R@10

Swipe to see all upstream methods →

ScanNet1500 · R@10 (%, higher is better)
Upstream	Initialization	+ PosePlaner
SuperGlue	62.4	68.9
LoFTR	25.0	52.9
RoMA	81.2	80.1
MASt3R	82.4	81.7
PoseDiffusion	9.0	38.8
RayDiffusion	33.4	50.5
DiffusionSfM	64.3	68.6

7Scenes · R@10

Swipe to see all upstream methods →

7Scenes · R@10 (%, higher is better)
Upstream	Initialization	+ PosePlaner
SuperGlue	65.9	69.7
LoFTR	34.2	56.8
RoMA	73.3	74.1
MASt3R	73.3	74.0
PoseDiffusion	12.2	44.6
RayDiffusion	32.7	51.4
DiffusionSfM	66.6	69.1

From two views to n views—without retraining. The same refiner, trained only on image pairs, refines pairwise poses within a long sequence; multi-view aggregation recovers the trajectory. Rotation AUC@5° (%) on the last 100 frames.

DatasetSequence length

ScanNet · N = 200 · R@5 (%, higher is better)
Upstream	Initialization	+ PosePlaner
CUT3R	23.3	40.5
TTT3R	46.2	50.7
WinT3R	37.7	45.9
Spann3R	32.4	45.6

ScanNet · N = 500 · R@5 (%, higher is better)
Upstream	Initialization	+ PosePlaner
CUT3R	16.5	37.3
TTT3R	36.1	47.1
WinT3R	28.3	42.5
Spann3R	20.0	38.8

ScanNet · N = 2000 · R@5

Swipe to see all upstream methods →

ScanNet · N = 2000 · R@5 (%, higher is better)
Upstream	Initialization	+ PosePlaner
CUT3R	17.3	37.3
TTT3R	27.5	48.2
WinT3R	19.2	27.9
Spann3R	19.4	37.1

7Scenes · N = 200 · R@5 (%, higher is better)
Upstream	Initialization	+ PosePlaner
CUT3R	22.0	35.1
TTT3R	44.1	44.5
WinT3R	40.0	43.9
Spann3R	34.5	40.8

7Scenes · N = 500 · R@5 (%, higher is better)
Upstream	Initialization	+ PosePlaner
CUT3R	16.2	35.1
TTT3R	36.0	41.9
WinT3R	44.2	46.3
Spann3R	35.7	42.1

7Scenes · N = 2000 · R@5

Swipe to see all upstream methods →

7Scenes · N = 2000 · R@5 (%, higher is better)
Upstream	Initialization	+ PosePlaner
CUT3R	12.0	28.1
TTT3R	38.4	41.7
WinT3R	41.6	42.9
Spann3R	39.6	43.7

Trajectory refinement

A ScanNet example: refined poses in red, initialization in blue, ground truth in black.

CUT3R trajectory on ScanNet Scene0757_00, comparing ground truth in black, raw estimates in blue, and refined estimates in red. — CUT3R

TTT3R trajectory on ScanNet Scene0757_00, with the refined tail in red closer to black ground truth. — TTT3R

WinT3R trajectory on ScanNet Scene0757_00, showing initialization in blue and refinement in red. — WinT3R

Zero-shot transfer, with no outdoor training. Both the plane predictor and pose refiner remain frozen on Oxford Spires and MatrixCity. Gains vary across settings, with some regressions, highlighting a limitation in outdoor generalization. Rotation AUC@5° (%).

DatasetSequence length

Oxford Spires · N = 200 · R@5 (%, higher is better)
Upstream	Initialization	+ PosePlaner
CUT3R	59.1	65.8
TTT3R	73.6	72.3

Oxford Spires · N = 500 · R@5

Swipe to see all upstream methods →

Oxford Spires · N = 500 · R@5 (%, higher is better)
Upstream	Initialization	+ PosePlaner
CUT3R	35.5	50.4
TTT3R	59.7	62.9

Oxford Spires · N = 2000 · R@5 (%, higher is better)
Upstream	Initialization	+ PosePlaner
CUT3R	34.0	53.3
TTT3R	37.2	51.3

MatrixCity · N = 200 · R@5 (%, higher is better)
Upstream	Initialization	+ PosePlaner
CUT3R	13.2	16.4
TTT3R	18.1	20.6

MatrixCity · N = 500 · R@5 (%, higher is better)
Upstream	Initialization	+ PosePlaner
CUT3R	12.4	15.1
TTT3R	10.5	14.4

MatrixCity · N = 2000 · R@5 (%, higher is better)
Upstream	Initialization	+ PosePlaner
CUT3R	9.9	15.5
TTT3R	12.7	17.2

Selected rotation metrics. Refinement gains vary with the initialization and setting; complete results are reported in the paper.

Citation

@misc{chen2026poseplanerdenoisingfeedforwardpose,
      title={PosePlaner: Denoising Any Feed-Forward Pose Predictor from Pairwise Planar Geometry}, 
      author={Yihang Chen and Bin Tan and Yujun Shen and Yinghao Xu and Nan Xue},
      year={2026},
      eprint={2610.34845},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.34845}, 
}