Project page · 2026

GeoWeaver: Accurate Long-Sequence 3D Reconstruction via Hierarchical Geometric Assembly

Tinghao Jiang^1,2, Sheng Tang², Shengzhe Wei², Juntong Fang³, Weiqi Zhang³, Junsheng Zhou³, Zesong Li²

¹ Shanghai Jiao Tong University ² Kosmo Research ³ School of Software, Tsinghua University

Paper coming soon Code BibTeX

GeoWeaver long-sequence reconstruction overview

GeoWeaver preserves accurate local geometric priors while assembling long RGB sequences into globally consistent cameras and point clouds.

Reconstruction videos

Explore indoor, outdoor, driving, and urban-scale reconstructions across four long-sequence benchmarks.

Courthouse

Tanks & Temples

Barn

Tanks & Temples

Meeting Room

Tanks & Temples

Virtual KITTI 2

VKITTI2

Keble College

Oxford Spires

Blenheim

Oxford Spires

Overview

Abstract

Long-sequence 3D reconstruction from RGB videos requires both accurate local geometry and globally consistent camera motion. Feed-forward models provide strong depth and pose predictions, but their memory cost prevents joint inference over long sequences. Chunk-wise processing improves scalability, yet independently predicted chunks often exhibit scale drift, pose errors, and point-cloud misalignment.

We present GeoWeaver, a unified framework comprising a Geometric Prior Model (GPM) and Test-Time Adaptation (TTA). The GPM predicts chunk-wise depth, confidence, and camera parameters as adjustable geometric priors. TTA then performs sequential initialization, global chunk-level Sim(3) alignment, and coarse-to-fine refinement of camera poses, affine depth corrections, and intrinsics. Dense correspondences provide adjacent, cross-chunk, and long-range constraints, while a robust CDF-style objective jointly optimizes weighted 2D reprojection and 3D consistency residuals.

Experiments across diverse long-sequence benchmarks demonstrate improved camera accuracy, global consistency, and point-cloud quality. Applying the same TTA procedure to different geometric prior models consistently improves their trajectory estimates, showing that GeoWeaver is not tied to a specific GPM.

Approach

Method overview

A feed-forward geometric prior model supplies reliable local predictions; hierarchical test-time adaptation turns them into a coherent global reconstruction.

GeoWeaver geometric prior model and hierarchical test-time adaptation pipeline — **GeoWeaver architecture.** GeoWeaver predicts chunk-wise depth, confidence, and cameras, then assembles them through sequential initialization, global chunk alignment, and frame-level refinement.

Chunk-wise geometric priors

A long RGB sequence is split into short chunks with one shared frame. The GPM predicts depth, confidence, camera pose, and intrinsics at bounded memory cost.

Dense correspondence routing

Confidence-aware dense matches connect temporal neighbors and long-range co-visible views, supplying constraints to each adaptation stage.

Hierarchical geometric assembly

Global Sim(3) alignment corrects chunk drift before joint refinement of frame poses, affine depth, and camera intrinsics.

Evaluation

Results

GeoWeaver produces coherent geometry, better aligned trajectories, and fewer discontinuities while using only one shared frame between adjacent chunks.

Oxford Spires Per-Scene Results

Relative rotation error (RRE, degrees) ↓ and absolute trajectory error (ATE, metres) ↓.

Method	Keble-04		Observatory-01		Blenheim-05		Christ Church-02		Average
Method	RRE ↓	ATE ↓	RRE ↓	ATE ↓	RRE ↓	ATE ↓	RRE ↓	ATE ↓	RRE ↓	ATE ↓
MBA Opt.	4.250	35.240	1.270	23.710	1.810	38.260	5.830	20.080	3.290	29.323
DA3 FF	14.442	22.664	7.087	8.365	2.123	1.538	16.193	33.992	9.961	16.640
VGGT-Long Chunk-FF	16.850	13.350	14.990	4.250	18.480	11.240	25.610	20.680	18.983	12.380
Scal3R Chunk-FF	7.830	2.130	5.270	1.570	7.430	2.560	6.830	16.000	6.840	5.565
LoGeR Chunk-FF	11.000	4.134	8.000	5.632	9.800	2.621	3.700	3.553	8.100	3.985
LingBot-Map Stream	2.090	8.870	0.970	6.960	1.040	4.570	1.240	3.880	1.335	6.070
AMB3R Hybrid	8.375	4.940	5.368	1.220	10.740	11.830	10.140	12.760	8.655	7.687
GeoWeaver Hybrid	0.432	1.430	0.360	1.820	0.330	1.430	1.180	12.810	0.575	4.372

Mip-NeRF 360 Visualization

Qualitative GeoWeaver point-cloud comparison on indoor and outdoor scenes — Representative comparisons on Observatory and the Mip-NeRF 360 room sequence.

Reference

BibTeX

Citation information will be updated when the paper is released.

@article{jiang2026geoweaver,
  title={GeoWeaver: Accurate Long-Sequence 3D Reconstruction via Hierarchical Geometric Assembly},
  author={Jiang, Tinghao and Tang, Sheng and Wei, Shengzhe and Fang, Juntong and Zhang, Weiqi and Zhou, Junsheng and Li, Zesong},
  year={2026}
}