VGGT-Bridge: Beyond Sequential Pose Graphs via Coarse-Stride Skip Edges

Sungjae Choi; Hanna Bae; Sunghyun Baek; Junmo Kim

cs.CVOct 5, 2026

VGGT-Bridge: Beyond Sequential Pose Graphs via Coarse-Stride Skip Edges

Authors: Sungjae Choi, Hanna Bae, Sunghyun Baek, Junmo Kim

Organizations: Korea Advanced Institute of Science and Technology, South Korea

Abstract

Feed-forward visual geometry transformers such as VGGT reconstruct dense 3D structure from images in a single forward pass, simplifying multi-view 3D reconstruction. However, their quadratic attention complexity makes them difficult to scale to long sequences with thousands of frames. Chunk-and-align frameworks address this by splitting a long sequence into overlapping chunks and stitching their local reconstructions into a pose graph. Yet existing methods connect only sequentially adjacent chunks, so small per-frame errors accumulate along the chain into large-scale drift. To move beyond sequential edges, we propose VGGT-Bridge, which adds long-range skip edges that directly constrain non-adjacent chunks without retraining. By running VGGT on sparsely sampled coarse chunks, each coarse chunk bridges distant fine chunks into a single direct constraint. We further turn VGGT's first-frame scale bias into a drift correction by feeding selected coarse chunks in reverse, and a loop-aware policy keeps this reversal compatible with existing loop closures. VGGT-Bridge reduces ATE by 28.3% on KITTI Odometry, 18.8% on Virtual KITTI, and 10.0% on Waymo Open over the SwiftVGGT baseline, achieving the best performance among all chunk-and-align methods.

Figures & tables

Figure 1 : Prior methods connect only sequentially adjacent chunks ( $\mathbf{S}_{1}\!\to\!\mathbf{S}_{2}\!\to\!\cdots$ ), so per-frame errors accumulate along the chain into large-scale drift. VGGT-Bridge inserts skip edges (pink arcs, e.g . , $\mathbf{S}_{1}\!\to\!\mathbf{S}_{3}$ ) that directly constrain non-adjacent chunks under the chunk-and-align pipeline. On KITTI Odometry Seq. 08, these long-range constraints lower ATE RMSE to $47.1$ m.

Figure 2 : Method overview. Gray, blue, and yellow boxes denote input frames, sparsely sampled coarse frames, and alignment overlaps, respectively. Standard chunk-and-align pipelines construct overlapping fine chunks and form sequential edges $\mathcal{E}_{\mathrm{seq}}$ (blue) between adjacent chunks, with optional loop edges $\mathcal{E}_{\mathrm{loop}}$ (purple) between revisited chunks. VGGT-Bridge additionally runs VGGT on sparse coarse chunks and composes two cross-scale $\mathrm{Sim}(3)$ alignments through each coarse chunk into a long-range skip edge $\mathcal{E}_{\mathrm{skip}}$ (rose) between non-adjacent fine chunks. Forward/reverse anchor markers show the selected processing direction of each coarse chunk, chosen by a loop-aware policy to complement the available loop closures. All edge sets are jointly optimized in a lightweight $\mathrm{Sim}(3)$ pose graph to recover globally consistent chunk poses.

Method	Calib.	00	01	02	03	04	05	06	07	08	09	10	Avg. ( ${\downarrow}$ )
# frames		4541	1101	4661	801	271	2761	1101	1101	4071	1591	1201
Length (m)		2453	5067	3724	561	394	2206	1233	650	3223	1705	920
Speed (m/fr)		0.82	2.23	1.09	0.70	1.45	0.80	1.12	0.59	0.79	1.07	0.77
Loop		✓	$\times$	✓	$\times$	$\times$	✓	$\checkmark$	$\checkmark$	$\times$	$\checkmark$	$\times$
ORB-SLAM2 [ 28 ]	✓	6.03	508.34	14.76	1.02	1.57	4.04	11.16	2.19	38.85	8.39	6.63	54.82
DPVO [ 40 ]	✓	113.21	12.69	123.40	2.09	0.68	58.96	54.78	19.26	115.90	75.10	13.63	53.61

Table 1 : Performance comparison on KITTI Odometry dataset. Bold / underline mark the best/second-best among the calibration-free methods. “Calib.” = uses camera intrinsics, “TL” = tracking lost, “OOM” = out of memory. Results with * are reproduced by us.

Method (segment ID prefix)	Calib.	1634	1838	3156	3461	3711	4058	4604	5200	6104	Avg. ( ${\downarrow}$ )
# frames		198	199	199	199	196	199	198	199	198
Length (m)		160	42	165	351	273	86	266	135	63
Speed (m/fr)		0.81	0.21	0.83	1.77	1.39	0.43	1.34	0.68	0.32
Traffic		Low	High	Low	Low	Med.	Low	Med.	Low	High
DROID-SLAM [ 39 ]	✓	3.705	0.301	0.447	8.653	9.320	7.621	4.170	TL	0.264	4.396
MASt3R-SLAM [ 29 ]	$\times$	4.500	0.556	1.833	12.544	8.601	1.412	5.428	7.910	1.195	5.560

Table 2 : Performance comparison on Waymo Open. Bold / underline mark the best/second-best among calibration-free methods. “Calib.” = uses camera intrinsics. Results with * are reproduced by us.

Method	Calib.	Scene01	Scene02	Scene06	Scene18	Scene20	Avg. ( ${\downarrow}$ )
DROID-SLAM [ 39 ]	✓	1.137	0.064	0.038	2.205	4.157	1.520
MASt3R-SLAM [ 29 ]	$\times$	TL					–
CUT3R [ 45 ]	$\times$	48.362	20.119	0.772	17.149	103.692	38.019
Fast3R [ 52 ]	$\times$	OOM					–
VGGT-Long [ 8 ]	$\times$	1.049	0.703	0.438	1.397	6.683	2.054
SwiftVGGT * [ 20 ]	$\times$	5.178	0.332	0.186	1.817	4.960	2.495

Table 3 : Performance comparison on Virtual KITTI. Bold / underline mark the best/second-best among calibration-free methods. “Calib.” = uses camera intrinsics, “TL” = tracking lost, “OOM” = out of memory (on at least one weather variant within the scene). Results with * are reproduced by us.

Figure 3 : Camera trajectory and dense 3D reconstruction comparison on KITTI Odometry.

Figure 4 : Scale bias on KITTI Odometry 00–10 (

n{=}11

sequences). Left: mean

\pm

std of the pairwise Umeyama

\log s

measured during chunk alignment, for sequential and coarse chunks. The arrow marks the shift

\Delta

from forward to reverse anchoring, with the percentage of edges where

\Delta{>}0

. Right: residual local scale error

\overline{|\log s|}

(bars) and the ATE RMSE (lines) of the final trajectory, for SwiftVGGT and our method. The scale error is measured on sliding

75

-frame windows against the ground truth.

	Translation ATE (m)											Mean
Configuration	00	01	02	03	04	05	06	07	08	09	10	Trans	Rot	Scale
# frames	4541	1101	4661	801	271	2761	1101	1101	4071	1591	1201
Loop	✓	–	✓	–	–	✓	✓	✓	–	✓	–
SwiftVGGT	9.92	114.96	55.45	10.38	5.48	12.50	8.38	3.50	69.74	30.92	25.55	31.53	13.38	0.157
+ coarse-stride chunk	13.00	69.68	53.43	9.26	4.44	12.30	7.28	4.68	72.43	24.06	25.45	26.91	$\mathbf{12.15}$	0.139
+ reverse coarse chunk	15.01	56.35	49.80	9.50	3.97	10.88	8.87	3.16	45.53	27.59	20.74	22.85	12.41	0.115

Table 4 : Component ablation on KITTI Odometry. Each row cumulatively adds one component on top of the previous. Trans and Rot are the translation (m) and rotation ( ∘ ) ATE, and Scale is the mean

|\log s|

of Fig. 4 (right).

Table 9

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter		KITTI Odometry	Virtual KITTI	Waymo Open
Fine chunk size $B$		75	75	60
Fine overlap $O$		30	30	24
Coarse ( $r{=}3$ )	$B^{3}$	35	35	35
Coarse ( $r{=}3$ )	$O^{3}$	14	14	14
Coarse ( $r{=}5$ )	$B^{5}$	25	25	25
Coarse ( $r{=}5$ )	$O^{5}$	10	10	10

Appendix

Table 7 : Hyperparameter settings.

				VGGT-Bridge
Dataset	# seq.	Mean $\log s$	SwiftVGGT	$\bar{w}{=}0.3$	$\bar{w}{=}0.7$	$\bar{w}{=}1.0$
KITTI Odometry	11	$-0.0370\pm 0.0225$	31.53	22.61	22.82	23.06
Virtual KITTI	30	$-0.0761\pm 0.0896$	2.495	2.027	1.853	1.790
Waymo Open	9	$-0.0470\pm 0.0341$	1.557	1.401	1.406	1.408
EuRoC MAV	11	$-0.0387\pm 0.0221$	1.722	1.698	1.632	1.638
7-Scenes	7	$-0.0167\pm 0.0169$	0.087	0.083	0.082	0.082

Appendix

Table 8 : Anchoring bias and skip-edge weight across five benchmarks. Mean

\log s

is the mean

\pm

std of the fine sequential edges over the sequences, as in Fig. 4 (left) of the main paper. The remaining columns report the mean translation ATE RMSE. Bold marks the best result in each row.

Method	00	01	02	03	04	05	06	07	08	09	10	Mean
SwiftVGGT [ 20 ]	9.92	114.96	55.45	10.38	5.48	12.50	8.38	3.50	69.74	30.92	25.55	31.53
VGGT-Bridge	13.74	56.34	48.70	9.33	4.00	11.02	6.93	3.15	47.15	26.24	22.05	22.61
$\Delta$ (%)	$+38.5$	$-51.0$	$-12.2$	$-10.1$	$-27.0$	$-11.8$	$-17.3$	$-10.0$	$-32.4$	$-15.1$	$-13.7$	$\mathbf{-28.3}$
VGGT-Long [ 8 ]	10.81	66.46	56.00	5.56	3.93	10.11	6.90	2.83	57.45	31.17	22.11	24.85
+ VGGT-Bridge	12.08	41.47	49.21	6.57	3.20	10.15	7.65	4.30	47.61	23.92	20.50	20.61
$\Delta$ (%)	$+11.7$	$-37.6$	$-12.1$	$+18.2$	$-18.6$	$+0.4$	$+10.9$	$+51.9$	$-17.1$	$-23.3$	$-7.3$	$\mathbf{-17.1}$

Appendix

Table 9 : Cross-framework transfer of VGGT-Bridge on VGGT-Long. (per-sequence translation ATE RMSE (m) on KITTI Odometry). The same skip-edge construction is added to two backbones, and

\Delta

is the change relative to each backbone.

	VGGT-Long [ 8 ]			SwiftVGGT [ 20 ]			VGGT-Bridge (ours)
Seq	Acc	Comp	CD	Acc	Comp	CD	Acc	Comp	CD
00	1.62	7.31	4.47	1.95	7.24	4.60	2.11	8.78	5.45
01	42.73	16.83	29.78	34.65	20.09	27.37	20.91	7.86	14.38
02	18.72	30.28	24.50	17.82	32.57	25.19	13.72	29.19	21.45
04	0.80	6.80	3.80	1.59	6.53	4.06	1.59	6.84	4.21
05	1.24	7.17	4.21	4.38	9.29	6.84	1.73	7.55	4.64

Appendix

Table 10 : Reconstruction quality on KITTI Odometry. Acc / Comp / CD (m), lower is better. Averaged over the ten sequences with available raw LiDAR scan, omitting Seq. 03. Best per metric per sequence in bold .

	VGGT-Long [ 8 ]			SwiftVGGT [ 20 ]			VGGT-Bridge (ours)
Segment	Acc	Comp	CD	Acc	Comp	CD	Acc	Comp	CD
1634	0.87	2.11	1.49	1.02	1.81	1.42	0.94	1.90	1.42
1838	0.45	5.87	3.16	1.38	5.85	3.62	1.35	5.44	3.39
3156	0.66	2.58	1.62	1.00	2.26	1.63	0.99	2.11	1.55
3461	0.93	4.47	2.70	1.38	5.53	3.45	1.36	5.50	3.43
3711	1.96	4.93	3.44	2.60	4.26	3.43	2.44	4.37	3.41

Appendix

Table 11 : Reconstruction quality on Waymo Open. Acc / Comp / CD (m), lower is better. Best per metric per segment in bold .

	Scene 01						Scene 02
Method	Clone	Fog	Morning	Overcast	Rain	Sunset	Clone	Fog	Morning	Overcast	Rain	Sunset
DROID-SLAM † [ 39 ]	1.03	1.87	0.99	1.01	0.78	1.15	0.10	0.04	0.05	0.05	0.04	0.11
MASt3R-SLAM [ 29 ]	TL
CUT3R [ 45 ]	43.30	62.19	50.61	38.73	51.55	43.79	23.77	9.95	28.41	24.64	7.96	25.97
Fast3R [ 52 ]	OOM
VGGT-Long [ 8 ]	0.76	0.87	0.93	0.67	1.80	1.26	0.72	0.71	0.72	0.68	0.69	0.69

Appendix

Table 12 : Per-scene, per-condition ATE RMSE (m) on Virtual KITTI. Best among the calibration-free methods in bold . “

\dagger

” uses camera intrinsics, all other methods are calibration-free. “TL” = tracking lost, “OOM” = out of memory. Results with * are reproduced by us.

	KITTI Seq. 07 (1,101 frames)			KITTI Seq. 08 (4,071 frames)
Stage	SwiftVGGT	Ours	$\Delta$	SwiftVGGT	Ours	$\Delta$
Loop detection	3.1	3.0	$-0.1$	10.2	10.1	$-0.1$
VGGT chunk inference	58.1	84.4	$+26.3$	197.6	307.9	$+110.3$
Loop-centric chunk inference	1.1	1.3	$+0.2$	0.0	0.0	$+0.0$
Loop-closure alignment	0.9	1.1	$+0.2$	0.0	0.0	$+0.0$
Sequential-edge composition	29.9	30.0	$+0.1$	76.0	76.0	$+0.0$

Appendix

Table 13 : Per-stage runtime breakdown on KITTI Odometry Seq. 07 and Seq. 08. Seq. 07 has one detected loop pair and activates the loop pipeline, while Seq. 08 has none detected. The loop pipeline is identical in both methods, so only the stages VGGT-Bridge modifies differ.

	GPU		Host
Method	Mean	Max	Mean	Max
SwiftVGGT [ 20 ]	10.30	14.49	15.74	29.90
VGGT-Bridge	10.30	14.49	19.81	38.90

Appendix

Table 14 : Peak memory on KITTI Odometry 00–10 (GiB), as the mean and the maximum over the eleven sequences. GPU is the peak memory reserved by PyTorch, and host is the peak resident memory.

	SwiftVGGT, $O$			SwiftVGGT	VGGT-Bridge
	$30$	$45$	$60$	$+B^{\prime}{=}90$	$O{=}30$
Frames to VGGT	$1.00\times$	$1.49\times$	$2.94\times$	$2.00\times$	$1.53\times$
Mean ATE (m)	31.53	32.61	34.70	32.91	22.61

Appendix

Table 15 : Compute-matched comparison on KITTI Odometry 00–10. Frames to VGGT is the total number of frames fed to VGGT, relative to SwiftVGGT.

Method	00	01	02	03	04	05	06	07	08	09	10	Avg. ( ${\downarrow}$ )
VGGT-Long [ 8 ]	9.87	111.06	37.56	4.89	3.75	9.09	7.47	4.02	62.86	47.48	25.49	29.41
SwiftVGGT [ 20 ]	8.17	102.53	36.49	8.12	4.88	11.94	8.88	5.01	64.68	44.13	26.18	29.18
VGGT-Bridge (ours)	11.61	61.21	38.18	7.50	3.91	9.92	7.71	4.63	49.29	35.19	25.84	23.18

Appendix

Table 16 : Full-transformation ATE on KITTI Odometry. Bold marks the best result.

Method (segment ID prefix)	1634	1838	3156	3461	3711	4058	4604	5200	6104	Avg. ( ${\downarrow}$ )
VGGT-Long [ 8 ]	3.09	2.54	2.34	4.01	4.04	3.14	2.87	3.41	2.33	3.08
SwiftVGGT [ 20 ]	3.11	2.45	2.35	2.72	4.08	3.11	2.84	2.83	2.21	2.85
VGGT-Bridge (ours)	3.20	2.01	2.36	2.28	4.53	2.94	3.03	2.66	2.22	2.80

Appendix

Table 17 : Full-transformation ATE on Waymo Open. Bold marks the best result.

Configuration	Mean ATE	$\Delta$ vs default
Default ( $\mathcal{R}{=}\{3,5\}$ , $W{=}60$ , $\tau{=}0.7$ )	22.61	—
Coarse stride set $\mathcal{R}$
$\mathcal{R}{=}\{3\}$ (drop stride 5)	24.52	$+1.91$
$\mathcal{R}{=}\{5\}$ (drop stride 3)	27.34	$+4.73$
$\mathcal{R}{=}\{3,5,7\}$ (add stride 7)	22.90	$+0.29$
Loop-density window $W$

Appendix

Table 18 : Hyperparameter sweeps on KITTI Odometry 00–10.

\Delta

is the change relative to the default VGGT-Bridge configuration (

\mathcal{R}{=}\{3,5\}

W{=}60

\tau{=}0.7

). Translation-only ATE RMSE (m) reported, as in the main tables.

Figure 5 : Additional camera trajectory and dense 3D reconstruction comparison on KITTI Odometry.

Figure 6 : Visualization of 3D reconstruction on Waymo Open. VGGT-Bridge reconstructions shown in RGB color (top) and colored by height, with red for low and blue for high (bottom).

Figure 7 : Visualization of 3D reconstruction on Virtual KITTI. VGGT-Bridge reconstructions shown in RGB color (top) and colored by height, with red for low and blue for high (bottom).

Explore similar work

Mamba-VGGT: Persistent Long-Sequence Video Geometry Grounded Transformer via External Sliding Window Mamba Memory
May 17, 2026Tianchen Deng, Zhenxiang Xiong, Nailin Wang +4Visual Geometry Grounded Transformer 3D Reconstruction
Shanghai Jiao Tong University · Nanyang Technological University · ETH Zurich +1
Attention Itself Could Retrieve.RetrieveVGGT: Training-Free Long Context Streaming 3D Reconstruction via Query-Key Similarity Retrieval
May 10, 2026Zichen Zou, Xiaosong Jia, Zuxuan Wu +1Visual Geometry Grounded Transformer 3D Reconstruction
Institute of Trustworthy Embodied AI (TEAI), Fudan University · 2Shanghai Key Laboratory of Multimodal Embodied AI
RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer
Jun 16, 2026Jinhao You, Shuo Lyu, Zhuohang Lyu +5Visual Geometry Grounded Transformer Self-Supervised Vision Transformers
University of Pennsylvania · University of California, Irvine · Nanyang Technological University +1

May 17, 2026, cs.CVJ/K move · Enter open · S save

Mamba-VGGT: Persistent Long-Sequence Video Geometry Grounded Transformer via External Sliding Window Mamba Memory

Tianchen Deng, Zhenxiang Xiong, Nailin Wang, Fangjinhua Wang+3

Shanghai Jiao Tong University · Nanyang Technological University · ETH Zurich+1

Visual Geometry Grounded Transformer 3D Reconstruction Transformer Architectures Bidirectional Long Short-Term Memory Catastrophic Forgetting

Open paper AskPDF arXiv

Visual Geometry Grounded Transformers (VGGT) have set new benchmarks in high-fidelity 3D scene reconstruction. However, as the sequence length increases, these models suffer from catastrophic geometric forgetting and accumulation drift, primarily due to the quadratic complexity of global attention which necessitates truncated temporal windows. To overcome the resulting geometric drift, we present Mamba-VGGT, an enhanced VGGT framework capable of persistent long-range reasoning. Our key contribution is a Sliding Window Mamba (SWM) memory module that maintains an explicit external memory token across temporal windows. This module leverages selective state-space modeling to distill and propagate global geometric priors, effectively bypassing the memory constraints of traditional transformers. To integrate these long-term temporal cues without disrupting the highly optimized spatial features of the pre-trained VGGT, we propose a Zero-Init Spatial Memory Injector. Utilizing zero-convolutional layers, this injector adaptively fuses persistent memory into the patch token stream, ensuring structural stability and seamless feature alignment. Extensive experiments demonstrate that our approach significantly outperforms existing VGGT-based methods in maintaining spatial consistency and reducing trajectory accumulation errors. Our work provides a scalable, linear-complexity solution for geometry-grounded world modeling in extensive 3D environments.

Abstract

Explore similar work

Mamba-VGGT: Persistent Long-Sequence Video Geometry Grounded Transformer via External Sliding Window Mamba Memory

Attention Itself Could Retrieve.RetrieveVGGT: Training-Free Long Context Streaming 3D Reconstruction via Query-Key Similarity Retrieval

RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer