Feed-forward 3D foundation models recover geometry at high fidelity, yet inference memory constraints keep them from processing long sequences natively. Current streaming systems work around this limit with sliding-window chunking, which introduces severe monocular scale ambiguity. Resolving this ambiguity requires Sim(3) registration that stays accurate and efficient under outlier contamination. We introduce Sim(3)-TurboReg, which reformulates the compatibility graph construction of TurboReg by replacing absolute distance constraints with a strict Tri-Scale Consistency (TSC) criterion. Our analysis shows that TSC suppresses both mismatching and coordinate-distortion outliers. We integrate the resulting solver into existing streaming systems and obtain high-fidelity, kilometer-scale reconstruction for feed-forward models. On controlled unknown-scale registration at 90% outliers, ours matches RANSAC accuracy with 92.5% lower median latency. In DA3-Streaming on KITTI, ours improves mean ATE RMSE by 22.2% and reduces mean end-to-end time by 16.6%. Code: https://github.com/ivan-zzj/Sim3-TurboReg
Figures & tables
Fig. 1: Sim(3)-TurboReg overview. Left: core registration on Stanford Bunny and Dragon at 60% and 90% outliers. At 90% outliers, TEASER++ fails while our method matches RANSAC accuracy at lower latency. Right: streaming integration with DA3-Streaming on KITTI-00 and TUM Floor-01, comparing IRLS and our backend. Sequence ATE RMSE decreases from 11.21 to 7.67 m (31.6%) on KITTI-00 and from 7.47 to 6.58 cm (11.9%) on Floor-01.
Fig. 2: Pipeline of Sim(3)-TurboReg. Putative matches are converted to tri-scale cues: an edge scale sij and two endpoint local scales si,sj . The TSC graph retains edges whose three scales agree ( Δij≤τ ), and 2nd-order scoring weights edges by shared-neighbor support. TurboClique extracts triplets, followed by a degeneracy filter for near-collinear configurations. Umeyama hypotheses are ranked by full-set inlier voting to obtain the Sim(3) estimate.
Fig. 3: Both outlier types in real streaming correspondences and their synthetic geometric models. (a) Type 1 mismatching outliers: paired points link different scene regions despite high descriptor similarity. (b) Type 2 coordinate-distortion outliers: pairs select the same landmark but undergo coherent displacement. Node scales use bilateral kNN ( k=6 ), τ=0.15 (illustrative settings; experiment defaults differ).
Outliers
Method
es / eR ( ∘ ) / et↓
CPU ms ↓ (Fail.)
60%
RANSAC-1K
0.0032 / 0.212 / 0.0056
0.119 (0)
RANSAC-1min
0.0032 / 0.212 / 0.0056
0.112 (0)
FGR-Sim3 †
0.1030 / 2.052 / 0.0994
0.141 (0)
TEASER++
0.0061 / 0.261 / 0.0076
0.914 (0)
Ours
0.0032 / 0.211 / 0.0056
0.703 (0)
70%
RANSAC-1K
0.0043 / 0.260 / 0.0071
0.260 (0)
TABLE I: Unknown-scale Stanford registration (320 trials/row, single-thread CPU). Mean errors and median time; et normalized. Errors as es / eR ( ∘ ) / et . † Sim(3) adaptation of FGR (our re-implementation). Best/second per metric: bold / underlined . Fail. counts trials with no estimate (out of 320). ∞ marks a mean that is undefined because the method returned no estimate on at least one trial; Table II uses the same convention.
Change vs. Ours
eR ( ∘ ) ↓
Fail. ↓
CPU ms ↓
TSC off
∞
910
0.648
No post-refinement
0.577
0
0.620
Weighted refinement
0.317
0
0.631
Union-median node scales
∞
2
0.647
Second-order scoring on
0.317
0
0.674
Degeneracy filter on
0.317
0
0.640
TABLE II: CPU ablations on the same 1,280 inputs as Table I , one factor changed relative to Ours. Mean rotation error eR over all trials and median CPU time. Fail. counts trials with no estimate (out of 1,280 ); a single failure makes the all-trial mean ∞ . Ours: histogram node scales, uniform post-refinement, TSC on, τ=0.1 , k=10 .
Type
Methods
Time [s]
Avg.
360
desk
desk2
floor
plant
room
rpy
teddy
xyz
Uncalibrated SLAM
Droid-SLAM Uncalibrated [ 8 ] †
–
0.163
0.202
0.032
0.091
0.064
0.045
0.918
0.056
0.045
0.012
Mast3r-SLAM Uncalibrated [ 28 ] †
–
0.060
0.070
0.035
0.055
0.056
0.035
0.118
0.041
0.114
0.020
Streaming Models
VGGT-Long [ 9 ]
868
0.0824
0.0632
0.0610
0.0601
0.1108
0.0616
0.1696
0.0374
0.1248
0.0534
Pi-Long [ 9 , 21 ]
1166
0.0808
0.1254
0.0580
0.0425
0.0957
0.0660
0.2052
0.0270
0.0863
0.0215
DA3-Streaming [ 20 ]
633
0.0492
0.0468
0.0202
0.0222
0.0747
0.0401
0.0948
0.0321
0.0907
0.0208
VGGT-Long + Sim(3)-TurboReg
492
0.0698
0.0651
0.0316
0.0412
0.1203
0.0490
0.1258
0.0339
0.1168
0.0443
TABLE III: TUM RGB-D ATE RMSE [m] ( ↓ ) and mean end-to-end time [s]. † =prior work. Best/second per group: bold / underlined .
Method
KITTI Odometry
TUM RGB-D
Avg.
00
01
02
03
04
05
06
07
08
09
10
Avg.
360
desk
desk2
floor
plant
room
rpy
teddy
xyz
VGGT-Long [ 9 ]
650
1785
299
1356
217
70
853
377
311
1109
452
322
868
789
577
456
861
1039
1313
737
1238
800
Pi-Long [ 9 , 21 ]
2337
5651
761
4481
649
263
3621
1261
1171
4525
2154
1168
1166
932
531
643
1433
1093
1759
1110
1852
1144
DA3-Streaming [ 20 ]
634
900
233
1597
228
59
869
490
474
1391
413
324
633
740
330
380
704
719
762
491
1096
472
VGGT-Long+Sim(3)-TurboReg
320
705
149
756
109
34
433
173
158
619
225
161
492
254
190
207
837
920
577
273
769
403
Pi-Long+Sim(3)-TurboReg
1419
4899
440
2074
276
92
2354
664
595
2700
773
738
695
543
364
312
1094
842
1145
340
913
701
TABLE IV: End-to-end processing time [s] (coarse; varies with system load). Best/second: bold / underlined .
Recent feed-forward geometry foundation models have demonstrated impressive generalization by recovering depth and poses in a single forward pass. However, these models are typically constrained by a global coordinate frame assumption. This dependency becomes a significant bottleneck for long-context and streaming reconstruction, as it forces the network to maintain an arbitrary temporal origin and handle translation magnitudes that grow unbounded over time. Our solution, which we call R3, employs relative regression. We employ a lightweight MLP to predict confidence-weighted relative constraints. These confidences serve as a unified anchor: weighting losses during training and guiding pose aggregation during inference. R3 supports both full-context offline reconstruction and causal, bounded-memory streaming. Our evaluation in both offline and streaming settings validates the effectiveness of our relative mechanism. Project page: https://kevinxu02.github.io/r3-site
Congrong Xu, Huachen Gao, Xingyu Chen +3
University of Michigan · Westlake University · NVIDIA
Recent 3D geometric foundation models, such as VGGT, provide robust feed-forward 3D reconstruction by directly predicting camera poses and 3D scene points from input images. However, their results remain inaccurate, and scaling them to long sequences or large unordered image sets typically requires chunk-wise processing, which can introduce drift and inconsistency. We present Glob3R, a global SfM-style reconstruction built on 3D foundation models. Our key idea is to explicitly optimize feed-forward geometric predictions. To this end, we augment a frozen Pi3X backbone with a lightweight dense matching head that predicts image warps between selected reference frames and neighboring views. These dense warps are converted into sparse but reliable multi-view feature tracks, which provide correspondence constraints for global optimization. We further introduce a keyframe-based sliding-window association strategy that propagates tracks and relative poses across overlapping windows, enabling scalable reconstruction. Finally, we perform global motion averaging and bundle adjustment to refine camera poses, reduce scale inconsistencies, and recover dense scene geometry. Extensive experiments on indoor, outdoor, large-scale driving, and unordered SfM benchmarks demonstrate that Glob3R achieves robust and accurate reconstruction. It consistently improves over feed-forward foundation-model baselines and recent scalable reconstruction methods, while being more robust than classical SfM pipelines. The refined poses also lead to higher-quality neural rendering, validating the benefit of combining foundation-model priors with global geometric optimization. Project page: https://junyuandeng.github.io/Glob3r
Junyuan Deng, Heng Li, Kejie Qiu +7
The Hong Kong University of Science and Technology · Tongyi Lab, Alibaba Group · Nanjing University +1
Feedforward foundation models have recently shown remarkable 3D reconstruction capabilities. However, existing models exhibit large tracking drift in long-context streaming reconstruction due to error accumulation. In this paper, we revisit loop closure with streaming reconstruction foundation models to enable accurate, drift-free, kilometer-scale reconstruction. Specifically, our method detects loop candidates through global descriptor retrieval, and constructs loop-conditioned windows to estimate the relative poses between looped frames. Given the observation that our adopted streaming reconstruction backbone produces a globally consistent scale, we optimize all frame poses on the SE(3) manifold with sequential and loop closure constraints, avoiding the pose graph optimization on the Sim(3) or higher-dimensional SL(4) manifolds employed in prior works. Extensive experiments show that our method reduces drift and produces consistent geometry on kilometer-scale sequences, significantly outperforming the state of the art. Code is available at https://github.com/MoyangLi00/CLoSeR.git.