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. This design preserves local geometric accuracy while correcting accumulated pose, scale, depth, and calibration errors. Experiments across diverse long-sequence benchmarks demonstrate improved camera accuracy, global consistency, and point-cloud quality. Ablations verify the contribution of each adaptation stage, and applying the same TTA procedure to different geometric prior models consistently improves their trajectory estimates, demonstrating that GeoWeaver is not tied to a specific GPM.
Figures & tables
Figure 1: Qualitative results across four benchmarks. Reconstructed point clouds and camera trajectories are shown for each scene.
Figure 2: Overview of GeoWeaver. The Geometric Prior Model predicts chunk-wise depth, confidence, and cameras, while Test-Time Adaptation performs sequential initialization, chunk alignment, and frame-level refinement using local and long-range correspondences.
Figure 3: Qualitative comparison on representative indoor and outdoor scenes. “Overlap” denotes the number of shared frames between adjacent 60-frame chunks. Compared with baselines using 30-frame overlap, GeoWeaver uses only one shared frame while producing more coherent geometry, better aligned trajectories, and fewer inter-chunk discontinuities.
Method
Depth
Camera
AbsRel ↓
SqRel ↓
AUC 5↑
AUC 30↑
DA3
0.086
0.088
53.2
77.6
π3
0.082
0.229
43.3
74.9
VGGT
0.125
0.602
43.2
68.7
Scal3R
0.147
0.170
34.9
67.0
LingBot-Map
0.114
0.177
29.0
64.3
Table 1: GPM evaluation across benchmarks. Camera AUC is reported at 5∘ and 30∘ . Best and second-best results are shown in bold and underlined, respectively.
Type
Method
T&T
Mip-NeRF 360
VKITTI 2
Oxford Spires
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
RRE ↓
ATE ↓
Opt.
MBA [ 46 ]
67.7
0.053
43.5
0.470
30.9
46.836
3.290
29.323
FF
DA3 [ 21 ]
63.9
0.358
66.4
0.037
16.9
38.867
9.961
16.640
Chunk-FF
VGGT-Long [ 7 ]
52.3
11.472
55.4
0.211
66.1
2.631
18.983
12.380
Chunk-FF
Scal3R [ 42 ]
58.1
0.338
37.8
0.194
67.4
0.608
6.840
5.565
Chunk-FF
LoGeR [ 44 ]
22.8
0.757
21.1
0.455
0.0
36.183
8.100
3.985
Table 2: Average trajectory results across four long-sequence benchmarks. T&T, Mip-NeRF 360, VKITTI 2, and Oxford Spires contain 6, 9, 5, and 4 evaluated scenes or sequences, respectively, spanning object-, room-, road-, and urban-scale reconstruction. AUC denotes AUC@3 ∘ and is reported in percentage; RRE is in degrees and ATE is in metres. Best and second-best results are shown in bold and underlined, respectively. Type abbreviations are Opt. (optimization-based), FF (feed-forward), Chunk-FF (chunk-wise feed-forward), Stream (streaming), and Hybrid.
Type
Method
Barn
Caterpillar
Courthouse
Ignatius
Meetingroom
Truck
Average
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
Opt.
MBA [ 46 ]
39.3
0.144
56.5
0.026
58.9
0.069
78.6
0.038
85.4
0.025
87.7
0.017
67.7
0.053
FF
DA3 [ 21 ]
69.5
0.039
63.7
0.023
32.4
2.000
67.4
0.047
76.9
0.021
73.3
0.021
63.9
0.358
Chunk-FF
VGGT-Long [ 7 ]
65.5
0.116
46.1
0.269
51.5
2.601
76.7
0.263
62.1
0.129
77.5
0.067
52.3
11.472
Chunk-FF
Scal3R [ 42 ]
59.9
0.053
44.6
0.158
43.8
1.656
59.9
0.076
71.3
0.039
69.4
0.046
58.1
0.338
Chunk-FF
LoGeR [ 44 ]
17.8
0.814
33.7
0.229
9.9
2.669
17.7
0.374
19.2
0.320
37.4
0.133
22.8
0.757
Table 3: Per-scene results on Tanks and Temples. AUC@3 ∘ is reported in percentage and ATE in metres. Best and second-best results are shown in bold and underlined, respectively.
Type
Method
Bicycle
Bonsai
Counter
Flowers
Garden
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
Opt.
MBA [ 46 ]
79.1
0.031
78.2
0.020
84.4
0.016
8.7
0.019
5.6
3.430
FF
DA3 [ 21 ]
63.1
0.054
76.4
0.027
77.1
0.016
72.9
0.028
64.4
0.039
Chunk-FF
VGGT-Long [ 7 ]
62.9
0.117
33.1
0.249
62.8
0.086
41.9
0.525
78.2
0.062
Chunk-FF
Scal3R [ 42 ]
42.9
0.120
27.5
0.170
63.7
0.033
5.1
0.570
65.2
0.059
Chunk-FF
LoGeR [ 44 ]
34.7
0.143
9.4
0.217
25.7
0.105
6.7
2.051
28.7
0.285
Table 4: Per-scene results on Mip-NeRF 360, split into two panels for readability. AUC@3 ∘ is reported in percentage and ATE in metres. Best and second-best results are shown in bold and underlined, respectively.
Type
Method
Scene01
Scene02
Scene06
Scene18
Scene20
Average
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
AUC ↑
ATE ↓
Opt.
MBA [ 46 ]
36.5
39.340
23.4
1.240
36.3
0.260
52.1
18.300
6.4
175.040
30.9
46.836
FF
DA3 [ 21 ]
3.9
22.726
45.3
0.151
26.1
0.075
4.6
12.061
4.4
159.321
16.9
38.867
Chunk-FF
VGGT-Long [ 7 ]
56.3
0.762
72.0
0.723
58.0
0.365
85.2
1.651
59.3
9.655
66.1
2.631
Chunk-FF
Scal3R [ 42 ]
81.3
0.376
74.5
0.038
42.4
0.031
83.3
0.165
55.6
2.430
67.4
0.608
Chunk-FF
LoGeR [ 44 ]
0.0
73.801
0.0
0.661
0.1
2.068
0.0
1.898
0.0
102.489
0.0
36.183
Table 5: Per-scene results on Virtual KITTI 2. AUC@3 ∘ is reported in percentage and ATE in metres. Best and second-best results are shown in bold and underlined, respectively.
Type
Method
keble-04
observatory-01
blenheim-05
christ-church-02
Average
RRE ↓
ATE ↓
RRE ↓
ATE ↓
RRE ↓
ATE ↓
RRE ↓
ATE ↓
RRE ↓
ATE ↓
Opt.
MBA [ 46 ]
4.250
35.240
1.270
23.710
1.810
38.260
5.830
20.080
3.290
29.323
FF
DA3 [ 21 ]
14.442
22.664
7.087
8.365
2.123
1.538
16.193
33.992
9.961
16.640
Chunk-FF
VGGT-Long [ 7 ]
16.850
13.350
14.990
4.250
18.480
11.240
25.610
20.680
18.983
12.380
Chunk-FF
Scal3R [ 42 ]
7.830
2.130
5.270
1.570
7.430
2.560
6.830
16.000
6.840
5.565
Chunk-FF
LoGeR [ 44 ]
11.000
4.134
8.000
5.632
9.800
2.621
3.700
3.553
8.100
3.985
Table 6: Per-scene results on Oxford Spires. RRE is reported in degrees and ATE in metres. Best and second-best results are shown in bold and underlined, respectively.
Figure 4: Convergence of GeoWeaver TTA. The objective decreases rapidly during Stage 2 chunk-level alignment and saturates during Stage 3 refinement.
Seq. Init.
Chunk Align.
Frame Refine
ATE ↓
AUC ↑
✓
–
–
0.674
24.4
✓
✓
–
0.075
43.8
✓
✓
✓
0.023
72.9
Table 7: Ablation of the three-stage Test-Time Adaptation procedure on Tanks and Temples. AUC denotes AUC@3 ∘ . The full configuration progressively introduces sequential initialization, chunk-level adaptation, and frame-level adaptation. Highlighted ablation results are shown in bold.
Prior Model
ATE ↓
AUC ↑
Scal3R [ 42 ]
0.338
58.1
Scal3R + TTA
0.033
63.6
DA3 [ 21 ]
0.358
63.9
DA3 + TTA
0.033
68.5
GeoWeaver-GPM + TTA
0.025
71.3
Table 8: Compatibility of Test-Time Adaptation with different Geometric Prior Models on Tanks and Temples. AUC denotes AUC@3 ∘ . The same TTA configuration improves multiple prior models without source-specific tuning. Highlighted ablation results are shown in bold.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Data domain
Datasets
Real indoor
ScanNet, ScanNet++, HM3D
Synthetic indoor
Hypersim, ASE
Procedural
MegaSynth
Multi-view / stereo
MVSSynth, Unreal4K, BlendedMVS
Dynamic / motion
DynamicStereo, TartanAir
Appendix
Table 9: Training datasets grouped by data domain.
Stage
Parameter
Value
Stage 1: sequential initialization
Confidence retention
50%
Stage 2: global chunk alignment
Learning rate
10−2
Maximum iterations
5,000
3D range / weight
0.8 / 2.0
Stage 3: coarse refinement
CDF range / bins
15 / 250
Smoothing bandwidth
2
Appendix
Table 10: Default hyperparameters of GeoWeaver TTA.
Figure 5: Residual distributions before and after optimization. The optimized 2D reprojection and 3D distance residuals concentrate closer to zero, producing steeper CDF curves and sharper PDF peaks.
Figure 6: Convergence curve of GeoWeaver TTA. The objective drops rapidly during Stage 2 chunk-level alignment and then saturates during Stage 3 refinement.
Figure 7: Additional qualitative results across diverse scenes. For each scene, we visualize the reconstructed point cloud together with the estimated camera poses.
We propose LoG-VGGT, a memory-efficient framework for long-sequence 3D reconstruction that balances local temporal modeling with global camera consistency. Instead of relying on full global attention, our method introduces cross-window attention at a small subset of transformer blocks, enabling effective information propagation across adjacent temporal windows while keeping memory usage bounded. To mitigate long-term pose drift, we further design a global camera consistency refinement module, where camera tokens interact with compact register tokens via cross-attention to enforce scene-level constraints across the entire sequence. This design enables joint optimization of camera representations and significantly improves long-horizon pose stability without incurring the high cost of sequence-wide attention. Extensive experiments demonstrate that LoG-VGGT achieves improved depth accuracy and robust camera pose estimation across multiple long-sequence benchmarks, while delivering competitive streaming reconstruction performance.
Jingke Zhou, Chenhang Ma, Zhizhou Zhong +6
Zhejiang University · Fudan University · Peking University +1
Maintaining global geometric consistency is a central challenge in long-sequence 3D reconstruction, with scale drift being the most critical failure mode. In chunk-based inference pipelines, the scale degree of freedom in sequential Sim(3) alignment is left unconstrained, causing estimation errors to compound multiplicatively and distort global trajectories and point cloud geometry. We present a scale-consistency enhancement framework built on a key insight: in structured environments such as driving scenes, geometric quantities arising from environmental regularity remain inherently invariant across temporal segments, and discrepancies in their per-chunk measurements directly expose inter-chunk scale drift. We propose Scene Geometric Invariant Anchoring (SGIA), which extracts dominant geometric invariants from each chunk's predicted point cloud via coarse-to-fine robust estimation and exploits their cross-chunk consistency to establish scale constraints independent of point cloud registration, explicitly degenerating 7-DoF Sim(3) alignment into 6-DoF rigid-body transformation and severing chain-wise scale error propagation at its source. We further introduce a lightweight test-time adaptation strategy that fine-tunes only normalization-layer parameters via multi-objective self-supervision, progressively improving intra-chunk predictions along the sequence. Both modules are plug-and-play and require no offline retraining. Experiments on multiple long-sequence benchmarks demonstrate state-of-the-art performance, reducing absolute trajectory error by up to 32% with significant gains in trajectory stability and reconstruction quality. Code: https://github.com/WZ-CS/VGGT-Align
Wei Zhang, Yihang Wu, Songhua Li +1
School of Computer Science Northwestern Polytechnical University Xi’an, China
Feed-forward 3D reconstruction provides an efficient paradigm for scene modeling from image sequences. Scaling these models to large monocular scenarios are constrained by excessive GPU memory footprint, degraded local geometry, and long-term trajectory drift. Existing chunk-based optimization strategies provide limited geometric constraints and fail to maintain global consistency over extended trajectories. We present a unified framework for stable and scalable feed-forward 3D reconstruction from long monocular sequences. Our approach builds on coarse-to-fine trajectory alignment augmented by lightweight geometric prior injection. Distilling monocular geometric cues into the feed-forward backbone via LoRA adaptation improves depth accuracy on fine structures while preserving inference efficiency. We introduce a hybrid-weight sparse ray-field optimization that leverages high-frequency geometric features to guide local point-cloud refinement and enforce consistent inter-frame ray constraints. Unlike prior chunk-based methods, this establishes strong cross-frame geometric coupling while maintaining scalability. Finally, an efficient trajectory stitching strategy with joint ray-error optimization explicitly reduces accumulated drift. Extensive experiments show that our approach achieves competitive trajectory accuracy compared with representative SLAM systems, while maintaining globally consistent 3D reconstruction in large-scale scenarios.
Enpeng Li, Yunzhou Zhang, Zhiyao Zhang +4
College of Information Science and Engineering, Northeastern University, China