BLASt3R: Bundle Adjustment of Any Image Set with Multi-View Matching and Monocular Priors

Vincent Leroy; Philippe Weinzaepfel; Lojze Zust; Yohann Cabon; Jérome Revaud

cs.CVSep 4, 2026

BLASt3R: Bundle Adjustment of Any Image Set with Multi-View Matching and Monocular Priors

Authors: Vincent Leroy, Philippe Weinzaepfel, Lojze Zust, Yohann Cabon, Jérome Revaud

Organizations: NAVER LABS Europe

Abstract

Recent hybrid Structure-from-Motion (SfM) systems combine the robustness of feed-forward 3D reconstruction with the accuracy of traditional bundle adjustment (BA) with pixel matching. They are usually the best performing methods however their scalability and usability remains limited since estimating dense correspondences between views is prohibitively costly, especially considering time constraints inherent to online applications like Visual SLAM (VSLAM). In this paper, we introduce a regularized BA framework that leverages a fast multi-view matcher and monocular priors for initialization and regularization. In contrast to existing systems, our unified approach seamlessly supports both online VSLAM and offline reconstruction from unordered image collections within the same optimization framework and sharing common hyperparameters for all tasks. Extensive experiments across both domains demonstrate improved performance and speed tradeoffs over traditional, feed-forward, and hybrid baselines. Notably for VSLAM, our uncalibrated method outperforms all previous calibrated approaches.

Figures & tables

Figure 1: Overview of BLASt3R , an integrated approach for dense 3D reconstruction that consists of: (a) a multi-view matching network that outputs dense flow from a query view to a set of memory views; (b) a monocular network that extracts adjustable pointmaps; and (c) a generalized bundle adjustment (BA) that efficiently ties these tracks and pointmaps together to form an accurate dense 3D reconstruction.

Figure 2: Exhaustive Coarse Matching (patch level) is performed with a transformer that updates the query tokens according to the current memory, followed by a dot product for exhaustive patch comparisons. The obtained scores are thresholded to get $\displaystyle\mathcal{P}$ , the estimated patch matches leveraged in the fine stage for pixelwise matching.

Figure 3: Sparse Fine Matching (pixel level) is performed only on the patch-level matches $\displaystyle\mathcal{P}$ from the coarse step. We update query tokens with a decoder, cross-attending matched memory tokens. For each pair of $\displaystyle\mathcal{P}$ we convert the token (patch) to pixel descriptors with an MLP followed by an unshuffle operation. We compare each pixel decriptor of the query image to all pixel descriptors of the memory by restricting the processing to $\displaystyle\mathcal{P}$ meaning we compute the similarity matrix with a block-sparse structure. We extract dense tracking with an argmax over patch pairs.

	Method	360	desk	desk2	floor	plant	room	rpy	teddy	xyz	Avg.
	ORB-SLAM3 [ 17 ]	$\displaystyle\times$	1.7	21.0	$\displaystyle\times$	3.4	$\displaystyle\times$	$\displaystyle\times$	$\displaystyle\times$	0.9	N/A
	DPV-SLAM [ 58 ]	11.2	1.8	2.9	5.7	2.1	33.0	3.0	8.4	1.0	7.6
	DPV-SLAM++ [ 58 ]	13.2	1.8	2.9	5.0	2.2	9.6	3.2	9.8	1.0	5.4
	DROID-SLAM [ 91 ]	11.1	1.8	4.2	2.1	1.6	4.9	2.6	4.8	1.2	3.8
	GlORIE-SLAM [ 121 ]	12.8	1.6	2.8	2.1	2.1	4.2	2.0	3.5	1.0	3.6
Calibrated	BLASt3R	5.2	1.4	2.4	2.0	1.3	5.1	2.0	2.6	0.9	2.5

Table 1: VSLAM on TUM-RGBD [ 86 ] fr1 subset. RMSE of absolute trajectory error (ATE) in cm, see Appendix C for extensive results.

Method / nb views	25	50	100	200	full
FlowMap [ 84 ]	2.0	5.1	12.6	30.7	27.5
ACE-Zero [ 12 ]	2.0	13.3	33.7	58.9	58.4
COLMAP [ 77 ]	13.1	29.6	48.1	62.0	-
VGGSfM [ 102 ]	57.9	64.3	62.5	40.8	0.0
DF-SfM [ 38 ]	48.0	63.0	67.8	23.3	52.6
MASt3R-SfM [ 28 ]	71.3	71.0	73.5	69.5	57.5

Table 4: Multi-view pose estimation on Tanks and Temples (average mAA@30 across different subsets of randomly sampled views per scene). OOM denotes Out Of Memory issues on an 80GB VRAM GPU.

Method	7-Scenes						ETH3D
	Acc. $\displaystyle\downarrow$		Comp. $\displaystyle\downarrow$		NC $\displaystyle\uparrow$		Acc. $\displaystyle\downarrow$		Comp. $\displaystyle\downarrow$		NC $\displaystyle\uparrow$
	Mean	Med.	Mean	Med.	Mean	Med.	Mean	Med.	Mean	Med.	Mean	Med.
VGGT	0.022	0.008	0.026	0.012	0.666	0.760	0.132	0.066	0.108	0.052	0.797	0.924
$\displaystyle\pi^{3}$	0.016	0.007	0.022	0.011	0.689	0.792	0.071	0.039	0.052	0.028	0.815	0.945
BLASt3R	0.021	0.008	0.036	0.012	0.674	0.768	0.069	0.026	0.055	0.017	0.881	0.973
BLASt3R nz=1	0.029	0.011	0.023	0.008	0.674	0.771	0.104	0.043	0.045	0.020	0.835	0.955

Table 6: 3D pointmap benchmarking following

\displaystyle\pi^{3}

’s protocol.

Method	NYUv2		KITTI		ETH3D		iBims-1		GSO		Sintel		DDAD		DIODE		Spring		HAMMER		Avg.
Method	Rel $\displaystyle\downarrow$	$\displaystyle\delta_{1}\uparrow$	Rel $\displaystyle\downarrow$	$\displaystyle\delta_{1}\uparrow$	Rel $\displaystyle\downarrow$	$\displaystyle\delta_{1}\uparrow$	Rel $\displaystyle\downarrow$	$\displaystyle\delta_{1}\uparrow$	Rel $\displaystyle\downarrow$	$\displaystyle\delta_{1}\uparrow$	Rel $\displaystyle\downarrow$	$\displaystyle\delta_{1}\uparrow$	Rel $\displaystyle\downarrow$	$\displaystyle\delta_{1}\uparrow$	Rel $\displaystyle\downarrow$	$\displaystyle\delta_{1}\uparrow$	Rel $\displaystyle\downarrow$	$\displaystyle\delta_{1}\uparrow$	Rel $\displaystyle\downarrow$	$\displaystyle\delta_{1}\uparrow$	Rel $\displaystyle\downarrow$	$\displaystyle\delta_{1}\uparrow$
ZoeDepth	4.76	97.3	5.59	95.1	7.27	94.2	5.85	95.7	2.54	99.9	21.8	69.2	14.2	80.1	7.80	90.9	24.3	66.6	6.65	95.7	10.1	88.5
MASt3R	4.67	96.7	5.79	95.1	4.64	97.0	4.62	95.6	2.85	99.4	21.3	70.3	12.5	83.4	5.79	94.1	27.4	62.8	4.21	96.8	9.38	89.1
DA V2	4.16	97.9	6.77	94.3	4.63	97.2	3.44	98.3	1.44	100	17.1	76.6	13.4	81.8	5.41	94.6	23.7	68.7	4.73	98.9	8.48	90.8
UniDepth V2	2.96	98.6	3.85	98.1	2.95	98.5	2.64	98.4	1.37	100	13.3	83.2	10.5	90.9	4.05	96.5	20.1	75.4	2.48	99.6	6.42	93.9
MoGe	2.92	98.6	3.94	98.0	2.69	99.2	2.74	97.9	0.94	100	13.0	83.2	8.40	92.1	3.16	97.5	4.34	96.4	3.00	98.3	4.51	96.1

Table 7: Monocular depth benchmarking following MoGe’s affine-invariant alignment protocol (nz=1). With nz=16, the alignment optimizes the best linear combination of adjustable depthmaps, i.e . this is an oracle performance bound.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A: Qualitative ablation study of the proposed multi-depth formulation (nz=16) versus vanilla affine alignment (nz=1) on relief-2 and meadow scenes from ETH3D [ 79 ] . Our method operates without prior knowledge about camera poses and calibration, and is aligned to ground-truth (GT scan) with a rigid alignment.

Figure B: Runtime comparison between VGGT-BA and BLASt3R on a system with a 80GB H100 GPU.

Panel of Figure 10 — Figure C: Additional qualitative examples. Top row: offline reconstruction examples. (Left) top-view of the delivery area scene from ETH3D MVS [ 79 ] , and Tanks&Temples Barn (right) . Bottom row: online VSLAM examples from TUM-RGBD [ 86 ] on the fr1 floor (left) and fr1 xyz (right) sequences. See accompanying video for animated visualizations.

Panel of Figure 11 — Figure D: Example of multiple depth channels generated by the monocular network. For each image, we show the main depth channel and the 14 first corrective channels. For the sake of visualization, corrective channels are orthogonalized using the Gram-Schmidt process to highlight corrections. See text for details.

Figure E: Example of multi-view dense matching directly output by our matcher network. Given the query image $\displaystyle I^{12}$ (top-left), dense optical flows are computed towards each of the memory images $\displaystyle\{I^{1},\ldots,I^{11}\}$ in a single forward pass. See text for details.

Figure F: Example of multi-view dense matching directly output by our matcher network. Given the query image $\displaystyle I^{30}$ (top-left), dense optical flows are computed towards each of the retrieved memory images in a single forward pass. See text for details.

Figure G: Example of multi-view dense matching directly output by our matcher network. Given the query image $\displaystyle I^{26}$ (top-left), dense optical flows are computed towards each of the retrieved memory images in a single forward pass. See text for details.

Structure from Motion

Structure from Motion estimates camera poses and 3D scene geometry from overlapping images by matching visual features across views.

See all 46 papers

VidMap: Exploiting Temporal Structure for Video-Based Structure-from-Motion
Jul 29, 2026Zador Pataki, Paul-Edouard Sarlin, Marc PollefeysMulti-View 3D Reconstruction Camera Pose Estimation
ETH Zurich · Google · Microsoft Spatial AI Lab
M^3: Dense Matching Meets Multi-View Foundation Models for Monocular Gaussian Splatting SLAM
Mar 17, 2026Kerui Ren, Guanghao Li, Changjian Jiang +7Visual SLAM Streaming 3D Reconstruction
Shanghai Jiao Tong University · Shanghai Artificial Intelligence Laboratory · Fudan University +5
Marginalized Bundle Adjustment: Multi-View Camera Pose from Monocular Depth Estimates
Feb 21, 2026Shengjie Zhu, Ahmed Abdelkader, Mark J. Matthews +2Camera Pose Estimation Structure from Motion
Google · Michigan State University

Jul 29, 2026J/K move · Enter open · S save

VidMap: Exploiting Temporal Structure for Video-Based Structure-from-Motion

Zador Pataki, Paul-Edouard Sarlin, Marc Pollefeys

ETH Zurich · Google · Microsoft Spatial AI Lab

Multi-View 3D Reconstruction Camera Pose Estimation Visual SLAM Structure from Motion Camera Motion Understanding

Open paper AskPDF arXiv

Accurately recovering the camera's calibration and metric poses for any unconstrained video would unlock large-scale training data for navigation and scene understanding. The dominant approaches to this problem are severely limited: Simultaneous Localization and Mapping (SLAM) is sensitive to initialization and transient failures due to its causal, incremental nature; it is often over-optimized for real-time operation and generally requires known camera calibration; while Structure-from-Motion (SfM) typically forgoes any image ordering, enabling optimal initialization and global optimization, but lacks robustness to visual symmetries and extreme motions. To bridge this gap, we introduce a system that combines the strong sequential constraints of SLAM with the flexibility and global optimization of offline SfM, enabling the metric reconstruction of arbitrary, long, uncalibrated videos. This system leverages recent advances in wide-baseline dense image matching, treats temporal ordering as a first-class citizen for reliable loop closure, and augments global optimization with metric monocular depth priors. As a result, thorough evaluations on diverse, challenging datasets that exhibit extreme motion and visual symmetries reveal that our approach is significantly more robust and accurate than both state-of-the-art SLAM and SfM, classical or learned, with given or unknown camera calibration. The code is publicly available at https://github.com/cvg/vidmap.

Abstract

Structure from Motion

Explore similar work

VidMap: Exploiting Temporal Structure for Video-Based Structure-from-Motion

M^3: Dense Matching Meets Multi-View Foundation Models for Monocular Gaussian Splatting SLAM

Marginalized Bundle Adjustment: Multi-View Camera Pose from Monocular Depth Estimates